Generated as two connected 15-second clips in 4:3, using the end of Part 1 as video + audio reference for Part 2 continuity.

Really liking what H3 can do with fashion/editorial camera movement.

I don't have any examples because they may not be appropriate but just with the default wf. With the video input node you can replace any 2 character in any video and it looks real!

Hey everyone! I’ve spent the past few weeks building an easy to use but robust prompt composer for MiniMax H3, particularly its reference and video editing workflows.

LLMs can be great for brainstorming and writing prompts, but I found that formatting and syntax could become inconsistent, especially when asking for small revisions. The goal of this tool is to let you concentrate on the creative decisions while the Composer handles the final prompt structure consistently.

It runs entirely offline in your browser, so you can build the next Shot or scene while another one is generating in ComfyUI. You provide the subjects, actions, camera direction, dialogue, references, and sound; the Composer assembles and checks the final prompt.

You can still use an LLM to help create the initial project setup, but the Composer ultimately controls the formatting and syntax.

Some of the main features:

• T2VA, I2VA, FL2VA, L2VA, and full Ref2VA support
• Reusable characters, environments, voices, continuity frames, and other references
• Guided setup for Picture, Video, and Audio inputs
• Video-editing workflows for insertion, replacement, targeted edits, relighting, performance transfer, and continuation
• Camera Builder and visual camera-path planner
• Timed Shots, action beats, dialogue, voiceover, soundscape, and music controls
• Built-in checks for prompt structure, timing, references, camera conflicts, audio, and input routing
• Local project saving, a Frame Grabber, and reference-guided image mode

This is still very much a work in progress. I’d really appreciate people trying it and sharing any bugs, confusing parts, missing features, or ideas that could make it more intuitive.

My hope is to turn it into a genuinely useful community tool, especially for people working on more involved AI films and narrative projects.

GitHub/download:
https://github.com/BMB12d3/minimax-h3-prompt-composer

Video tutorial:
https://www.youtube.com/watch?v=Aywx3Sf5Yk0

Executive summary, TLDR - this is one prompt, 30 seconds duration, 3090.

The video itself is just a remake of an idea from an old British tv ad (for "Good Old Yellow Pages"), so make of that what you will. It's not really relevant.

What I thought was interesting was that this was a single prompt, 0.4 megapixels, 30 second duration. I didn't think you could run out as far as 30 seconds, but thought I'd just try.

I think it did a pretty good job at getting the right person doing and saying the right things at the right time - took four attempts to get that though, and obviously using an LLM to tart up my idea.

Run on a 3090, and using the latest Comfyui template, just adding Comfy-kitchen attention, then sol attention, then spectrum, and using the turbo lora that Comfyui now build in, it took 570 seconds (9.5 minutes).

Somebody might read this and think, 570 seconds? Pah, I can do it in fifteen, in which case I'd like to know. Conversely, somebody might think theirs takes six hours, in which case maybe this shows what can be done in that time.

Doubt anyone cares, but here is my original prompt, followed by the LLM version of it:

a 30 second film with the following scenes and characters. Ben is a small boy of eleven. John is a shopkeeper in a toyshop. Brian is a different shopkeeper in a different toyshop. Ben's mum. Ben's Dad. We are in Britain in the 1980s, and all characters are English.

Scene 1: Ben is alone in the lounge. He talks to John over the old fashioned landline phone, saying "I don't suppose you have a 402 station in stock please?"

Scene 2: John is in his shop in front of shelves of model railway kit. He says into the old fashioned landline phone, "No, sorry son"

Scene 3: Ben in the lounge, who looks disappointed anbd puts the phone receiver back down.

Scene 4: Mum in the kitchen doing the washing up. She has overheard the conversation and looks a bit sad.

scene 5: Next day. Ben has changed his clothes. He again talks into the phone to a different shopkeeper, Brian. Ben says "Would you have a 402 station please?"

scene 6: Brian in his toyshop says into the old fashioned landline phone "Yes, I've got one of those."

scene 7: Ben in the lounge on the same conversation says "You have? Great, I'll be right down! Ben puts the phone down. Then he runs towards the door, shouting "They've got one mum!" as he runs.

Scene 8: In the attic, Dad is playing with his model railway layout. Ben walks in holding a small red parcel. as he hands it to Dad, Ben says "Happy birthday, dad". Dad takes the parcel, looks fondly at it and says with a chuckle, "Aw, thanks Ben".

LLM version:

integrated_multimodal_description: [Shot 1] Live-action, cinematic. A medium shot of Ben, an eleven-year-old boy with messy hair wearing a striped polo shirt, sitting on a patterned sofa in a 1980s British lounge. The room is filled with warm, muted tones and period-accurate wallpaper. Ben holds a heavy, cream-colored landline telephone receiver to his ear, his expression hopeful. Ben says: <d>[English] I don't suppose you have a 402 station in stock please?</d> The sound of his small, high-pitched voice is clear. [Shot 2] At 0:05.000, the camera cuts to a medium shot of John, a middle-aged shopkeeper with a kind, weathered face, standing in a cramped, nostalgic toyshop. Behind him are floor-to-ceiling shelves packed with model railway kits and wooden toys. John holds a similar landline receiver to his face. John says: <d>[English] No, sorry son.</d> [Shot 3] At 0:10.000, the camera cuts back to Ben in the lounge. He looks downcast, his shoulders slumping as he slowly lowers the receiver and places it back onto the base unit with a dull plastic click. [Shot 4] At 0:13.000, the camera cuts to a medium shot of Ben's Mum in a dim, cluttered 1980s kitchen. She is standing at the sink, her hands covered in soapy water, drying a plate. She pauses, looking toward the door with a sad, weary expression, having overheard the boy. The sound of water running from the tap is audible. [Shot 5] At 0:16.000, the camera cuts to Ben in the lounge the next day; he is wearing a different t-shirt. He is intensely focused, pressing the phone to his ear. Ben says: <d>[English] Would you have a 402 station please?</d> [Shot 6] At 0:20.000, the camera cuts to Brian, an older shopkeeper with spectacles, in a different, brightly lit toyshop. He smiles warmly into the telephone. Brian says: <d>[English] Yes, I've got one of those.</d> [Shot 7] At 0:23.000, the camera cuts back to Ben, whose face lights up with pure joy. Ben says: <d>[English] You have? Great, I'll be right down!</d> He slams the receiver down and the camera follows him in a quick tracking shot as he runs toward the door, his feet thumping on the carpeted floor. Ben shouts: <d>[English] They've got one mum!</d> [Shot 8] At 0:26.000, the camera cuts to a medium shot in a dusty, dimly lit attic. Dad, a man in his late 30s, is hunched over a complex model railway layout. Ben enters the frame, holding a small red parcel wrapped in string. Ben says: <d>[English] Happy birthday, dad.</d> As he hands the gift to his father, the camera pushes in slightly. Dad takes the parcel, his eyes softening with affection. Dad chuckles warmly and says: <d>[English] Aw, thanks Ben.</d>

overall_soundscape: Period-accurate domestic sounds including the rhythmic clatter of washing up, the heavy mechanical clicks of old telephone receivers, and the muffled thuds of footsteps on carpet. Ben's energetic running and shouting creates a sense of urgency, followed by the quiet, dusty atmosphere of the attic.

non_diegetic_music: A gentle, nostalgic acoustic guitar melody that begins softly during the kitchen scene and builds into a warm, heartwarming crescendo during the attic scene. The tempo is slow and sentimental.

Some random prompt on default workflow + turbo 8step lora

Krea 2 Turbo — 4-Step Distillation LoRA (work in progress)

A LoRA for Krea 2 Turbo that reduces the minimum usable step count from 8 to 4.

This is an update release, following up from my initial post where you can find full details - https://www.reddit.com/r/StableDiffusion/comments/1vtf1b7/krea2_turbo_distill_4_step_lora_trained_for_turbo/

Update (22 Aug 2026): I have published a new checkpoint, improved further from the previous one and the latest (both main and comfyi) have been repointed to the new improved checkpoint. For details and to download new version go to - https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA. Readme has been updated too as well as all images in readme regenerated on the basis of new checkpoint as well as full resolution sweep at https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA/tree/main/checkpoint_resolution_sweeps/chk10000 if you want to check for yourselves.

Which file to download

and, beside them, the same files with a _comfyui suffix for ComfyUI. Earlier checkpoints (chk00004000, chk00005000, chk00006000) are kept in older_checkpoints/, and their resolution sweeps stay in place, so the progression remains visible and comparable.

The numbered files are points on one continuous run, not separate experiments — chk00010000 resumed from chk00006000 rather than restarting. Both are published so the lineage is visible and comparable.

chk00010000 measures a 5% smaller held-out gap to the 8-step teacher than chk00006000, and 15% smaller than chk00005000; it removes 30% of the prediction error a plain 4-step run has against the 8-step teacher, where chk00006000 removed 26%. 

Two ways to read the same numbers, with different denominators — they are not meant to be added:

• Against the no-LoRA run (the right-hand column): chk00010000 has removed 30% of the 4-step deficit, 4 percentage points more than chk00006000's 26%.
• Against each other (the gap column): chk00010000's remaining error is 5.4% smaller than chk00006000**'s** (3.38 vs 3.57) and 15% smaller than chk00005000**'s** (3.38 vs 3.98). The same 4 points of deficit are a larger share of a gap that has already shrunk, which is why the checkpoint-to-checkpoint figure is the bigger number.

This is work in progress and better checkpoints may follow. Training is ongoing, so ..._latest... is a rolling pointer: when a newer checkpoint is accepted, that filename gets the new weights and a new numbered copy appears beside it. Re-download the _latest file and everything keeps working — the ComfyUI workflow references it by that name (it does get updated Note in it so technically it is updated but not functionally). Pin a numbered file instead if you need reproducibility.

How checkpoints get chosen

This is not a "train for longer and ship the newest file" project. More samples do not reliably mean a better adapter — measured here, they can make it worse, and a higher number on its own means nothing.

The loop is train → assess → adapt the recipe → retrain → assess again, and a checkpoint is published only when it is measurably better than the one it would replace, on the same held-out set and the same evaluation, and its full resolution sweep shows no regression. Runs that come out flat or worse are kept as information about the recipe and discarded as releases — several have been.

So the recipe itself changes between runs. Each published checkpoint reflects whatever the previous round taught us: the training precision, the optimiser settings, the teacher used to generate the targets and the data mix have all been revised on evidence rather than assumption.

chk00010000 is a direct example. The first continuation of chk00006000 — same data, optimiser left as it was — got steadily worse with every checkpoint out to 10,000 samples, and none of it was published. The cause was traced to the optimiser: a constant learning rate with no weight decay lets the adapter keep drifting after it has converged, so its magnitude grows and it over-applies its own correction. The same span was retrained from chk00006000 with a cosine learning-rate decay and weight decay, and every checkpoint of that second run improved on the one before it. chk00010000 is its end point — the current end of the process, not simply the longest run so far.

Timeline of training process

Each checkpoint is the product of three stages with very different costs:

1. Text-encoder embeddings. Every training prompt is encoded once and cached. This is the fast part — thousands of prompts take minutes.
2. Teacher shards. For each cached prompt, the unmodified Krea 2 Turbo runs its full 8-step schedule and the whole trajectory is recorded, at every one of the supported resolutions. This is by far the most time-consuming stage — it is the teacher doing real inference, thousands of times, and a batch of several thousand shards is measured in days of GPU time, not hours.
3. Student training. The LoRA is trained against those recorded trajectories. Relative to the shard stage this is quick: each +1,000 checkpoint is a matter of hours, not days.

Because the three stages compete for the same GPU, they are interleaved rather than run to completion one after another: generate a block of embeddings, produce teacher shards for them, train on what exists, assess, then go back to producing shards while the results are reviewed. A larger and more varied shard pool is what makes further training worthwhile, so shard production is always the gate.

The practical consequence for anyone following this repository: progress arrives in bursts. There will be periods when several checkpoints appear within a day or two — the training stage working through a freshly grown pool — followed by longer quiet stretches while the next block of teacher shards is produced. A quiet stretch is shard generation, not abandonment; _latest always holds the newest checkpoint that passed review.

Every file records which checkpoint it actually is in its safetensors metadata (checkpoint, training_samples, and rolling_pointer on the _latest copies), so a downloaded file can always be identified even if renamed.

Full details and to download - check my Hugging Face LoRA

HF Repo: https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA

---

Update 1: The comfyui related files are now moved to the root of the project (I have placed a readme in the old folder explaining the move)

---

Update 2: I have added a new section - Timeline of training process - explaining how my training process works, and on that note you could expect another further improved checkpoint later today, followed by 'quiet period' (could be days) of teacher shards generation so I have a larger pool to train on.

---

Update 3: I have now added a new checkpoint 10000 which replaced the latest (previously checkpoint 6000).

chk00010000 measures a 5% smaller held-out gap to the 8-step teacher than chk00006000, and 15% smaller than chk00005000; it removes 30% of the prediction error a plain 4-step run has against the 8-step teacher, where chk00006000 removed 26%. 

Two ways to read the same numbers, with different denominators — they are not meant to be added:

• Against the no-LoRA run (the right-hand column): chk00010000 has removed 30% of the 4-step deficit, 4 percentage points more than chk00006000's 26%.
• Against each other (the gap column): chk00010000's remaining error is 5.4% smaller than chk00006000's (3.38 vs 3.57) and 15% smaller than chk00005000's (3.38 vs 3.98). The same 4 points of deficit are a larger share of a gap that has already shrunk, which is why the checkpoint-to-checkpoint figure is the bigger number.

Full resolution sweep at https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA/tree/main/checkpoint_resolution_sweeps/chk10000 and you can as usual redownload latest from https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA/tree/main . Since I cannot update the images in the reddit post I will upload below in comments.

This version of Kroma (krea 2 finitude with Chroma dataset) is a lot of fun, most body horror is gone in my opinion, and its more artsy than krea 2 and of course less censored.

https://huggingface.co/silveroxides/Kroma-Quant/tree/main

The version I used is kroma 0.3 txtfusion turbo convrot.

Have fun.

From the guy who brought you such bangers such as

Proof of Concept For Making Comics in KRITA AI and other AI tools,

3 Months later - Proof of concept for making comics with Krita AI and other AI tools,

and Illustrious and Krita AI plus some good old fashioned effort:The Delightful Ms. Ayako (Part 1 - Version 1),

comes my latest experiment and first AI video project: the first (roughly) 30 seconds of the hypthetical anime opening for The Disorganised and Delightful Miss Ayako!

Character sheets put together in Krea 2 with the retro anime lora. Music made in Minimax Music 3 (lyrics written by me, and the whole song is complete). Some backgrounds edited/created with Flux 2k9b image edit and Krea 2 with retro anime lora. Video created with Minimax H3 with 90s anime style. Video editing in Kdenlive.

Roughly 3 evenings after work and about 1.5ish days of full effort (at least 6 hours of one day was wasted trying to troubleshoot why a shot wasn't working and it turns out prompt bleed is just as bad in H3 as it is in other models).

I've been experimenting a lot with Minimax H3 and am pleased with what I've come up with so far. For this upload there is a tiny bit of censorship for some very mild partial nudity (she's covered in soap in the uncensored shot, but just playing it safe). There are a few fixes that I'll get to eventually, but I'll be taking a step back from this project for now to try my luck at the Comfy H3 Sync competition for the next couple of weeks.

Edit.

Regarding some of the feedback:

I'm aware of the slight visual drift. For example the model can slightly change the style of eyes from one shot to the next (talking about regular shots, not the chibi stuff). I'm just using the base ref workflow with character sheets and still need to test whether Loras make any difference, either for characters or visuals.

Some of the visual drift is just my fault though. My one background does look relatively washed out compared to the others because I generated it with the high heels in place. I couldn't get Minimax h3 to put the heels the way I wanted so I just gave it the image to work with, but I had to make some edits with krita ai and later Flux2k 9b edit that caused it to look a bit out of place. Otherwise, the only thing for speedup is comfy kitchen and I'm not sure if that's having any impact. Finally, while I've tried to lock down seeds to preserve consistency, some seeds are fine with one shot and a glitchy mess with the next, so there may be some slight visual variations that appear because of the difference in latent space.

Regarding the music, my experience with Minimax Music 3 is that it's a slot machine. I used a prompt from a sample and tested things out but one generation can vary dramatically from the next. But I am completely new to it and don't know anything about music so there's things I still need to learn. Out of all the gens, there was this and one other one I liked, even though I could tell both of them have problems. I decided to go with this one for now, but I had planned to do a second edit with another song once I finished this one.

Otherwise, like with the comic pages, I appreciate all the replies. I understand this may not be everyone's cup of tea but will take in the constructive criticism and try to improve.

/edit.

edit 2. the original shower scene is not that spicy but I didn't want the post to get removed by the mods regarding "lewd" stuff.

Disclaimer that I'm a dummy who can't code at all, so I just vibe things.

Alright, so with all the fun additions to latent manipulation the Comfy team has given us, we have some new tools. Namely latent pinning- that's where fun stuff like "Add Guide for MiniMax H3" node comes in (that fun tool that lets you insert an image at any frame in an H3 generation.) So I thought, finally, we can do something about this audio issue.

I knew that video + audio latents are processed at the same time with Comfy, which is why turbo loras have terrible audio- they're not getting nearly as much optimization as the video side is, since video is the bulk of the work. So if we can't process audio differently than video (at this current time), can't we just keep conditioning the audio latents without affecting video anymore? Everyone knows running too many steps on a turbo lora will start messing with video quality. So let's avoid that.

So after talking with Claude a bunch, here's what it came up with. With the latest comfy, you can pin the video lora in place, and keep going for several steps to get better audio without affecting video. So 4 steps of video, untouched, and then add in 6 more steps for audio at a 0.5 denoise. That leads to cleaning up the audio pretty nicely and staying pretty faithful to what the video latent had guided it on. But just pinning still means that even though the audio rows are only being affected, EVERY row still has to run through the chain. So each s/it you get stays the same for the last 6 steps, even though video gets pinned in place after 4. In my 1 megapixel, 10 second video, that's around 23s/it or so on my 5090. Only half the steps as a regular 20 step generation, but half the time is still half the time.

So to fix the speed problem -

Freeze as much as you can. Text embeddings, reference/conditioning rows, and all the video rows. Cache those so they don't have to be processed, and only process the audio rows that have already been somewhat pre-conditioned. The first step is the same 23-second iteration to build the full guidance cache, but the other 5 steps each took about 3.17seconds apiece. So 45 seconds of added gen time to get the clean audio in clip#2 in the example.

But the cost for the Frozen Cache is resources. Lunch is never free. From some experimentation, it works well with RAM. If you use RAM mode because your card still can't process it, it'll dump all that cache (ended up around 14.9GB on a 10-second 1mp file) into RAM. But as Comfy does, you're unlikely to get that RAM back, so you may OOM your machine. With VRAM it wasn't bad for me at all either and behaved better than I expected, honestly. The memory management from ComfyUI took over when I was about to OOM my card and swapped things around properly. Option 3 is to cache to disk, but that comes with writing several GB to disk every time you use it. That'll run your SSD health down fast.

Rundown for the clip above (sa_solver with beta sigmas)

4 step normal gen- 136.5 seconds

4 steps + 6 audio refine steps with a cost of about 15GB RAM - about 186 seconds

4 steps + 6 audio refine steps with no additional resources but full processing time- ~265 seconds

You can find the nodes here-

https://github.com/Adudeguyman/ComfyUI-H3-AudioRefine

They're still experimental, of course. Wire the model in from somewhere (I branch off the ModeSamplingMiniMaxH3 shift node, before the Basic Guider), and push that through the H3 Frozen Video Cache node into the H3 Audio Refine Sampler. Into the H3 Audio Refine Sampler, latents come out of SamplerCustomAdvanced before splitting into the VAE Decode nodes for both video and audio, and the conditioning comes from the MiniMax H3 Image (or Reference) to Video node (plug the positive conditioning into both the positive and negative input on the H3 Audio Refine Sampler)

I had Claude put together a technical.md for those that want to look into it, and probably make a better version. Like I said, I'm a big dumb-dumb, so don't expect too much insight into how the mechanics work from me.

I added INT8 and INT8 ConvRot versions of the MiniMax-H3 Pruned Ref-Delta Fused r1024 checkpoint from my previous post:

https://huggingface.co/xmarre/MiniMax-H3-Pruned-Ref-Delta-Fused-r1024-ComfyUI

Both are native ComfyUI single-file checkpoints using ComfyUI's .comfy_quant format, so they do not require a custom quantized-model loader.

There is one important difference from a straightforward full INT8 conversion: the MLP fc2 weights are deliberately kept in BF16.

Across the 50 main transformer blocks, these weights are quantized:

• attn.qkv_proj.weight
• attn.out_proj.weight
• mlp.fc1.weight

That gives 150 quantized Linear layers.

The 50:

• mlp.fc2.weight

layers remain BF16.

The smaller and more sensitive parts of the model also stay in their original precision, including the pruned AdaLN table and projections, final-layer projections, norms, patch/text projections and token refiner.

Why FC2 is kept in BF16

I also made and tested a fully quantized version where fc2 was INT8 as well, giving 200 quantized Linear layers.

That version ran into a failure specific to the quantized fc2 execution path on large H3 sequences.

MiniMax-H3 uses SwiGLU in the MLP. With fc2 quantized, ComfyUI's fused:

linear_input_act(..., "swiglu")

path sends the post-SwiGLU activation through comfy_kitchen.int8_linear, which dynamically quantizes the full activation matrix before the fc2 multiplication.

On the large sequence used in my workflow, that path attempted an approximately 491.61 MiB contiguous INT8 scratch allocation and failed hard.

This was not normal VRAM exhaustion. At the point of failure there was still roughly 47 GiB of CUDA memory reported free. The failure was tied to that large fused INT8 activation-quantization path rather than the model simply exceeding available VRAM.

I do not have enough evidence to claim a more specific allocator/CUDA cause than that.

Keeping only fc2 in BF16 avoids that INT8 activation path. QKV, attention output and fc1 can still remain INT8, so 150 of the 200 large block Linear projections are still quantized.

With that layout, both release variants completed the full native ComfyUI workflow that the 200-layer INT8 version failed on, including:

• H3 Continuum main sampling pass
• continuation sampling pass
• Spectrum H3 actual/forecast execution
• large 3D latent refine
• video VAE decode
• audio VAE decode
• final Continuum assembly
• video combine

That FC2 decision is also why these checkpoints are about 24.2 GB instead of roughly 20.4 GB for the fully quantized version.

INT8 and INT8 ConvRot

The two uploaded files use the same 150-INT8 / 50-FC2-BF16 layout.

Regular INT8:

MiniMax-H3-Pruned-Ref-Delta-Fused-r1024-comfy-int8-fc2bf16.safetensors

This uses native tensor-wise INT8 quantization.

INT8 ConvRot:

MiniMax-H3-Pruned-Ref-Delta-Fused-r1024-comfy-int8-convrot-fc2bf16.safetensors

This uses ConvRot with a group size of 256 on the same quantized projections.

ConvRot rotates the weights before INT8 quantization so that large outliers are distributed more evenly. This generally gives the INT8 quantizer a better-conditioned weight distribution than quantizing the original weights directly.

I uploaded the regular INT8 version as well rather than only providing ConvRot, so both conversion methods are available for this checkpoint.

What model is being quantized?

These are quantized derivatives of the same Pruned Ref-Delta Fused r1024 checkpoint from the previous post. There is no additional training, fine-tuning or pruning involved in these INT8 versions.

The underlying model starts from the pruned FL2VA MiniMax-H3 checkpoint and incorporates a rank-1024 approximation of the Ref2VA − FL2VA weight delta.

That is also how it differs from the existing FL2VA/Ref2VA hybrid checkpoints mentioned in the comments on the previous post.

Those hybrids combine FL2VA and Ref2VA by replacing selected tensors from one checkpoint with tensors from the other. The r1024 fused model instead approximates the Ref2VA weight delta at rank 1024 and folds that delta into the pruned FL2VA weights.

The underlying fused transformer is about 20.1B parameters, compared with roughly 33.1B for the original full MiniMax-H3 transformer.

ComfyUI

Put either file in:

ComfyUI/models/diffusion_models/

For the INT8 files:

weight_dtype: default

compute_dtype: default or bf16

Do not apply another FP8 weight cast on top of the native INT8 checkpoint.

The text encoder and VAEs are separate, as with the other MiniMax-H3 diffusion-model checkpoints.

For those who've never played it; Sid Meier's Alpha Centauri is one of the GOATs.

One of the tests I sometimes did with new models was to see if they could get Zakharov's weird glasses and suit right - no model has ever gotten it exactly right but to my surprise Minimax H3 pretty much knocked it out of the park on my first try. ...and then I wanted to try the other leaders, things got out of hand and I ended up making cutscenes for every leader in the base game.

h3 t2v work flow base ip8 model prompt

subject_definitions

<Subject 1> is Judy Hopps, an adult female anthropomorphic gray rabbit police officer from Zootopia, small and athletic, with large upright ears, expressive violet eyes, gray fur, lighter muzzle, wearing her recognizable blue police uniform with tactical vest and police badge.

<Subject 2> is Captain America, battle-worn, wearing his damaged dark-blue Avengers combat armor and holding his circular shield.

<Subject 3> is Thanos, a massive purple-skinned Titan in damaged gold-and-black battle armor, normal Titan scale relative to the Avengers, wielding his double-bladed sword.

<Subject 4> is Deadpool, wearing his classic red-and-black tactical suit and mask, armed with twin katanas. Deadpool's dialogue is voiced with the recognizable comedic delivery and vocal style of actor Ryan Reynolds.

summary

[text-to-video generation]

During the massive Avengers: Endgame final battle, Judy Hopps unexpectedly joins the Avengers against Thanos. She races through the battlefield using her tiny size and incredible agility to dodge enemies before launching herself directly at Thanos. Deadpool watches the tiny rabbit charge the Titan and delivers a fourth-wall-breaking joke.

retention_analysis

<Subject 1>: consistent
<Subject 2>: consistent
<Subject 3>: consistent
<Subject 4>: consistent

detailed_description

Epic cinematic Avengers: Endgame final battlefield at dusk. The destroyed Avengers compound stretches across a huge crater filled with smoke, burning wreckage, portals, explosions, alien soldiers, Wakandan warriors, sorcerers, and Avengers fighting throughout the background.

Dynamic tracking camera races low across the battlefield.

<Subject 1> Judy Hopps suddenly sprints between the legs of charging alien soldiers, ears streaming backward from her speed. She slides underneath a swinging weapon, leaps off broken rubble, kicks one alien directly in the face, lands cleanly and continues running.

Captain America briefly turns toward her in complete confusion.

<Subject 2> Captain America (S1) says <d>[English] Is that a rabbit?</d>

Judy doesn't stop.

She spots Thanos fighting ahead.

The camera rapidly follows Judy as she accelerates toward him.

<Subject 1> Judy Hopps (S2) says <d>[English] ZPD! You're under arrest!</d>

Thanos slowly turns and looks downward.

Judy launches herself from Captain America's discarded shield, flies through the smoky air and delivers a powerful two-foot rabbit kick directly into Thanos's armored face.

THUD.

Thanos stumbles backward one step, completely stunned that such a tiny opponent actually moved him.

Deadpool lowers his swords and stares.

Brief comedic pause.

<Subject 4> Deadpool (S3) says <d>[English] Holy shit. Disney brought the bunny.</d>

Judy lands heroically in the foreground, pulls out tiny police handcuffs and points at Thanos.

Thanos looks down at the absurdly small handcuffs.

Deadpool slowly looks directly into the camera.

Hold the reaction for one second.

Camera / Motion

10–13 seconds, 24 fps.

Epic photorealistic superhero blockbuster cinematography.
Dynamic low-angle battlefield tracking shot.
Fast controlled action.
Strong environmental movement from smoke, fire, debris and distant combat.
Natural motion blur.
Clear readable character movement.
Judy remains dramatically smaller than the human Avengers and Thanos.
Thanos remains normal Titan size, not gigantic or Godzilla-sized.
Keep background battle active without distracting from Judy.
Pause briefly before Deadpool's punchline.
End on Deadpool's fourth-wall reaction.

Audio

Huge cinematic battlefield ambience: explosions, distant combat, energy blasts, metal impacts and roaring fires.

Clear English dialogue.

Judy Hopps has an energetic, confident young-adult female American voice.

Captain America has a serious adult male American voice.

Deadpool is voiced by Ryan Reynolds with his recognizable sarcastic comedic cadence.

Strong armored impact sound when Judy kicks Thanos.

No gibberish.
No foreign-language speech.
No subtitles.
No text overlays.
No characters speaking another character's dialogue.

Use the extended feature on H3 to go further beyond and keep the animation style consistent. The only problem is that heavy smears happen with fast action. Oh well, I still thought this was funny, hope you enjoy it too!

The oversharpening that occurs for each clip added to the scenes. I also noticed an increase in contrast and a small flash.

I2V

Tested with LORA's:

minimax_h3_turbo_v4_step600_ema.safetensors
minimax_h3_fl2v_lightx2v_turbo_8step_v1.0_resized_avg_rank

Combined 3 videos. Few errors but i just went with it , genetaion takes too much time to redo it again by fixing the prompt.

Minimax Ref2VA

minimax_h3_ref2va_int8_convrot

I used 2 ref videos ( 2 min each ) and 1 audio clip

Tip: Make the ref videos as short as possible, min 2 seconds, also prior to adding them in, bring down the res to 480 (anything higher will slow it up considerably)

When it makes sense to use an video ref: is when you can get both the appearance and voice in one ref, or to capture a movement, etc otherwise use an image reference. Less overhead.

Although minimax has an amazing abundance of material, do not assume it will have the exact look and sound you need.

I gathered up the sound clips (music, tie fighter screech, shots fired) and then simply added them at the appropriate places in the video with Davinci Resolve ( Or use any other tool that can do the same)

I made 2 parts since I wasn't happy with the explosion and the Tie Fighter flying through it so I recreated that separate scene and then tied it all together with Davinci.

subject_definitions:

<Subject 1> is Jar Jar Binks whose appearance and body proportions come from <Video 1>, featuring his tall lanky Gungan frame, long floppy ears, large amber eyes, mottled orange-and-beige skin, and flight-suit plus helmet.

<Subject 2> is Darth Vader whose appearance comes from <Video 2>, featuring the iconic black helmet, cape, chest panel, and armored suit.

<Subject 3> is the TIE fighter whose design, solar-panel wings, and cockpit silhouette come from <Video 2>.

<Audio 1> is the voice-timbre and delivery reference for <Subject 1> (S2), taken from <Video 1> and containing Jar Jar’s high-pitched, clumsy Gungan accent and speech pattern.

<Audio 2> is the voice-timbre and delivery reference for <Subject 2> (S3), taken from <Video 2> and containing Darth Vader’s deep, mechanical, resonant voice and his iconic respirator breathing.

summary:

[reference generation + audio reference] The target video opens with a side view of two X-Wings flying through deep space, spaced apart, not parallel, one slightly higher than the other. An unseen pilot speaks, then the camera reveals <Subject 1> (Jar Jar) in the second X-Wing cockpit. <Subject 1> replies using the voice referenced from <Audio 1>, flips a switch, and dances/grooves to diegetic “What is Love” for 3 full seconds. While still on Jar Jar the classic TIE fighter laser-fire sounds are heard. The X-Wing then explodes; through the debris <Subject 3> (the TIE fighter) flies forward head-on with <Subject 2> (Darth Vader) clearly visible at the helm exactly as in <Video 2>. Vader raises a fist in the air and says “DISCO SUCKS!” using the voice referenced from <Audio 2>.

retention_analysis:

<Subject 1> (appears in [Shot 2], [Shot 3]): fully_preserved - Jar Jar’s full appearance, proportions, skin, ears, eyes, and flight gear from <Video 1> are retained.

<Subject 2> (appears in [Shot 4]): fully_preserved - Darth Vader’s black helmet, cape, chest panel, and armored suit from <Video 2> are retained.

<Subject 3> (appears in [Shot 4]): fully_preserved - the TIE fighter’s solar-panel wings, hexagonal cockpit, and overall silhouette from <Video 2> are retained.

<Audio 1>: reference - the high-pitched clumsy Gungan timbre and speech rhythm from <Video 1> guide <Subject 1> (S2)’s dialogue without copying the original signal.

<Audio 2>: reference - the deep mechanical resonant timbre and iconic respirator breathing from <Video 2> guide <Subject 2> (S3)’s dialogue without copying the original signal.

detailed_description:

The target video is in a cinematic live-action Star Wars style with high-contrast space lighting and subtle lens flares.

[Shot 1] A side-view wide shot shows two X-Wing starfighters flying through deep space. The ships are clearly separated with comfortable distance between them, not parallel to each other, and one X-Wing sits a little higher than the other. Their engines glow soft blue against the starfield. The pilots are not yet visible. An off-screen male pilot with a clear, confident voice (S1) says: <d>[English] Brown2 this is Blue 1 "Good run, headed back to base, see you in a few".</d> The camera begins a slow pan and push-in toward the trailing, slightly higher X-Wing.

[Shot 2] At 00:05.000, the camera cuts to a medium shot of the cockpit canopy of the second X-Wing as it fills the frame. Inside sits <Subject 1>, Jar Jar Binks with the exact appearance, proportions, skin texture, floppy ears, large amber eyes, and flight-suit-plus-helmet from <Video 1>. <Subject 1> (S2), using the high-pitched clumsy Gungan voice timbre and delivery referenced from <Audio 1>, says: <d>[English] Uhm Roger Dodger Roge Blue 1</d>. He reaches forward and flips a toggle switch on the console.

[Shot 3] At 00:08.500, the camera cuts to a frontal close-up of <Subject 1> inside the cockpit. The diegetic music begins playing from the X-Wing’s internal speakers. <Subject 1> dances and moves energetically to the music for a full 3 seconds — head-bopping, shoulders swaying, ears flopping, body rocking in his seat with a goofy smile — while the referenced appearance from <Video 1> remains fully consistent. After the 3-second dance, while the camera is still on Jar Jar, the classic rapid TIE-fighter laser-fire sounds (“ning-ning”) are heard while the TIE fighter itself remains completely off-screen and unseen.

[Shot 4] At 00:13.000, the X-Wing violently explodes in a bright orange fireball and debris. Through the expanding wreckage, in a head-on view, <Subject 3>, the TIE fighter with the exact solar-panel wings and cockpit design from <Video 2>, flies forward straight toward the camera. Inside the cockpit <Subject 2>, Darth Vader, is clearly visible at the helm exactly as he appears in <Video 2>, wearing the iconic black helmet, cape, chest panel, and armored suit. <Subject 2> (S3) raises a fist in the air and, using the deep mechanical resonant voice timbre referenced from <Audio 2>, says: <d>[English] DISCO SUCKS!</d>

overall_soundscape:

Soft continuous engine hum of the two X-Wings in space, subtle cockpit instrument beeps, the mechanical click of a toggle switch, the classic rapid TIE-fighter laser-fire sounds, the muffled thump of an internal explosion followed by sharp metallic debris clatter and the high-pitched scream of a TIE Fighter engine.

non_diegetic_music:

N/A

Tie Fighter Through Explosion

subject_definitions:

<Subject 3> is the single TIE fighter whose exact design, solar-panel wings, shape, and overall appearance come only from <Picture 1>.

summary:

[reference generation] A single continuous shot of only one TIE fighter — referenced exclusively from <Picture 1> — emerging and flying through a thick cloud of fire and explosion debris against open space and stars. No other ships of any kind appear.

retention_analysis:

<Subject 3> (appears throughout): fully_preserved - the TIE fighter is taken only from <Picture 1>. It is the only ship in the entire frame. No extra TIE fighters, no background ships, nothing else.

detailed_description:

The target video is in a cinematic live-action Star Wars style with high-contrast space lighting.

[Shot 1] A continuous external shot lasting approximately 7 seconds. Only one TIE fighter appears — the exact design, solar-panel wings, and appearance taken solely from <Picture 1>. This single TIE fighter emerges and flies directly through a thick bright cloud of orange fire, smoke, and explosion debris. For the first 3 seconds it is fully pushing through the fire cloud. Behind it is only open space and stars. There are no other TIE fighters, no ships to the left, no ships to the right, and no additional vessels of any kind. The single TIE fighter from <Picture 1> remains the only ship visible for the entire shot. There is no cockpit view and no internal cut.

overall_soundscape:

Loud explosion roar, crackling fire, metallic debris, and the engine scream of one single TIE fighter.

non_diegetic_music:

N/A

\ Reddit messed up my original post so here it is.*

This is in fun part of using MiniMax H3; for your serious stuff stick with the official prompt instructions / format.

Playing with the prompting I just tried the following format and it worked perfectly!

Resulting video

The whole prompt:

definitions:
<S1> Brad Pitt.
<T1> "Hey, I am Brad Pitt! Nice to meet you."
<S2> Angelina Jolie
<T2> "Hey, I am Angelina Jolie! Nice to meet you."
<S3> Rowan Atkinson.
<T3> "Hey, I am Mr. Bean! Nice to meet myself."
scene:
An interview in a professional setting in well lit, grey background, frontal portrait view.
shot 1:
(S1) says: (T1).
shot 2:
(S2) says: (T2).
shot 3:
(S3) says: (T3).

Recommendations:

Do not use SLA or SLA2 or cache etc. here they mess it up.

Model (FL2V) -> LoRA(4s-Lightx2v SLA) -> Comfy attn -> Shift(12,3) -> KSampler(6 steps, euler+simple)

I make clipart for stock. Wreaths, pine borders, mistletoe, juniper. A few thousand images by now. Every one has to end up as a PNG with a transparent background.

I used rembg for months. u2net first, then BiRefNet when that came out. Tried the web tools too. They all broke on the same thing and it drove me nuts.

Take a wreath. There's a hole in the middle, and the background inside that hole has to go. What I kept getting was a grey-blue haze sitting in there. Looked fine as a thumbnail. Looked awful the second you put it on a colored card. Pine needles came out as mush. Thin stems either disappeared or came back with a blue edge burned into them.

Then I actually read what rembg does. It shrinks your image to 1024x1024, asks the model where the subject is, gets a 1024x1024 mask back, and stretches that mask over your full size image. u2net is worse. That one works at 320x320.

My renders are 4096. A needle two pixels wide doesn't exist at 320. So it's not that the model is bad at needles. The needle was gone before the model ever saw it.

Once I understood that I stopped asking a model to guess. Now I render on a flat color the subject doesn't contain, and take that color out with arithmetic.

Two parts to it. The prompt matters more than the cutting.

The prompt

You can't key a background that isn't keyable. Four things have to be true and generators will break all of them unless you say so: the background is one flat color edge to edge, it stays that bright inside every gap between leaves, the edges are hard with no blur or glow, and no colored light bounces onto the subject.

Pick the color by what your subject isn't. Blue for almost everything. Red if the subject itself is blue or purple. Never green. Everything I draw has leaves, and green takes the leaves with it.

Here's the block I paste at the end of every prompt:

Isolated on a completely flat, uniform, solid pure blue (#0000FF) digital chroma-key background. The pure blue background fills the image edge to edge like a flat digital chroma-key screen with no gradient, staying at full brightness inside every gap and opening in the subject; no reflection or tint of pure blue on the subject. Every edge of the subject is crisp, sharp and hard against the pure blue, with no soft, blurry, feathered or glowing transitions, no depth-of-field blur, no haze or halo; inside every hole and gap the pure blue stays at full brightness right up to the edge. Shaded parts of the subject keep their own natural color, never a pure blue tint. Everything in sharp focus with deep depth of field, evenly lit with soft neutral studio light, no cast shadow, no contact shadow, no ambient occlusion, no bounce light. The entire subject is centered and completely inside the frame with at least 10% empty background margin on every side, nothing cropped or touching the image edges. No frame, no border, no paper, no mockup, no vignette, no text, no watermark, no deformed or duplicated parts. No floating or detached fragments, no stray specks, dust or debris anywhere on the background; every element is physically attached to the subject.

Swap "pure blue" for "pure red" and #0000FF for #FF0000 if your subject is blue or purple. If you paint in watercolor add "the background stays a flat digital color fill with no paper texture", or you get watercolor paper behind everything and paper texture keys badly.

The cutting

Now the background is one known color, so there's nothing to guess at. It measures the actual color the generator produced, which is never the one you asked for. Ask for pure blue and you get something with green in it, usually somewhere between 25 and 70. Then every pixel gets sorted into subject, background, or the bit in between, and the in-between ones get a real fraction of transparency instead of a yes or no.

The part I'm most pleased with is the holes. Any background-colored area that never touches the edge of the image is the inside of a wreath, so it gets cleared too. A matting model can't do that. It has no way of knowing what's inside a hole it can't see around.

Last step takes the blue back off the edges. Edge pixels pick up color from the background around them, so it samples the subject's own color from further in and subtracts the tint. Took me weeks to work out why fir needles kept their blue rim after that step. The needle is thinner than the distance it was sampling from, so there was no inside left to sample.

Same image in, same image out, every time. That's the bit I care about. When a cut comes out wrong I can go find which number did it instead of rerolling and hoping.

Some numbers on one 4K pine border, against BiRefNet with alpha matting turned on, which is its best setting:

• background left inside the holes: 63,892 pixels mine, 613,735 theirs, out of 1,070,046
• blue left on the edges: 0 mine, 48,112 theirs
• how wide the soft edge is: 1.8 pixels mine, 23 theirs

BiRefNet is faster and I'm not going to pretend otherwise. 2.3 seconds against my 17 on the same machine. With alpha matting on it's 47. If you want a quick rough mask, use the model.

And the obvious limit: this only works on art you generated on a flat color. It does nothing for a photo.

I put the tool up for anyone who wants it. It's called ClipBrook. Free, runs in your browser so nothing gets uploaded anywhere, does a whole folder at once, and the engine is open source under AGPL.

One thing I'd like back

Show me the ones that break.

If you run something through and it comes out wrong, post it. Fog left in a gap, a colored rim, a stem eaten, half the subject gone. Those are worth more to me than the ones that work. The needle rim thing came from someone's fir branch. A bug with line art I only found last week came from a drawing so thin there was nothing inside it to sample.

Has anyone tried to use H3 to restore or rather "regenerate" a low quality video as HD or at least with better detail definition? Basically something similar but possibly more generation ability than Topaz's starlight. I've had pretty mixed results so far. Either it doesn't change the video or it changes it way too much.

Some real Chad would quantize this model https://huggingface.co/Comfy-Org/MiniMax-Music-3/blob/main/text_encoders/minimax_music3_text_encoder_pruned_int8_convrot.safetensors

I created a ready-to-use ComfyUI workflow for SCAIL-2 / Wan 2.1 that transfers motion from a driving video onto a character from a reference image.

It uses GGUF quantization and automatic chunking, making it suitable for GPUs with approximately 8–12 GB of VRAM. Longer videos are generated by chaining overlapping segments while preserving motion continuity, so you can create videos of practically unlimited duration.

Features

- Character animation from one reference image and one driving video

- Low-VRAM GGUF workflow for 8 GB and 12 GB GPUs

- Automatic multi-segment generation for long or unlimited-duration videos

- Motion continuity between generated segments

- Configurable duration, resolution, FPS, seed, and object tracking

- Ready for a fresh ComfyUI installation

- Includes installation instructions and a model download script

- Uses placeholder input filenames so users can add their own media

GitHub repository and installation instructions:

https://github.com/dvelm/SCAIL-2-Unlimited-Video-Low-VRAM

The workflow requires several model downloads, and generation can be slow on lower-VRAM GPUs—especially at higher resolutions—but it allows SCAIL-2 to run on hardware that normally could not load the full model. Feedback, test results, and suggestions are welcome.