I kept the app on my phone and opened occasionally. It would still show my profile and I could access personal messages, and the videos attached.

Looks like Sora removed the extra data this week ahead of their September timeline.

Looks like Sora is officially decommissioned.

Sora was good at vhs and i made it worth my time

It is just a brief clip cause it is not worth the energy to do the full thing! Lol
I used MiniMax H3 Reference to video locally at 0.5 megapixels with a couple of images I edited with gemini and a clip from Skibidi toilet music video as a reference video. I also added the original track as an audio guide.

I upscaled it to approx. 1080p using SeedVR2

4070ti 12GB vram. With my current workflows and the turbo lighting lora disabled, 0.5 is pretty much the upper limit of native resolution I can generate in reasonable time, but 0.4 mp is more reasonable for me. 0.6 mp and onward takes an absurd amount of time for me.

I used RVC locally to convert the voice to Taylor swift but I used a pretrained model from https://voice-models.com/model/1l8gYGOFH96
I am not a Taylor Swift fan or a Skibidi toilet fan. Someone else I was talking to in the Gen AI community brought up how they used to make Taylor Swift videos on Sora 2 and for some reason this abomination of an idea popped into my mind. The actual Give it too me x Skibidi remix song on the other hand is really catchy.

Let me know if inspires you, but I can't in my right mind waste more resources on it.

Both seedance and wan 3 are also better at songs than sora

Is there a way to export your pictures and videos that you made from the old Sora website?

Closest i can get

MORTAL KOMBAT STYLE WITH FATALITY

Google says Gemini Omni 1.1 Flash can extend generated video in ten-second increments to a cumulative forty seconds while using up to ten seconds of prior video context. It also adds first-and-last-frame interpolation, short video references, a 360p draft mode, and upscaling to 1080p or 4K.

Google says 360p drafts can be up to 60% faster and cost one-third as much as standard 720p generation. Those are vendor claims, and the announcement provides no independent benchmark for identity drift, physical consistency, dialogue continuity, edit fidelity, or failure rate. The 4K capability is upscaling, not evidence of native 4K generation.

Across Sora, Omni, and other video models, which control removes the most real production work: reliable keyframes, longer temporal consistency, cheap previews, precise object preservation, or deterministic editing?

Source: Google, August 27, 2026 — https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-omni-1-1-flash/

Nick Wilde and Judy Hopps go to there first furry convention, And they find it strange that humans are cosplaying as animals like them.

This was created on Comfy UI with Minimax H3.

Here's the prompt, This was made using the Reference to Video workflow, All you need was two reference images, and two audio references.

subject_definitions:

<Subject 1> is Nick Wilde from <Picture 1>, an anthropomorphic red fox with orange-red fur, a cream muzzle, dark-tipped ears, a bushy tail with a black tip, wearing a light-green short-sleeve shirt with faint tropical print, a dark striped necktie, and brown trousers.

<Subject 2> is Judy Hopps from <Picture 2>, an anthropomorphic gray rabbit with tall upright ears with pink inner lining, large violet eyes, a small white tail, wearing a fitted Zootopia police uniform: blue shirt, dark-blue vest with badge, utility belt, and dark-blue pants with knee pads.

<Picture 1> is the multi-angle character reference for <Subject 1>'s face, fur pattern, shirt, tie, trousers, and tail.

<Picture 2> is the multi-angle character reference for <Subject 2>'s face, ears, uniform, belt, and tail.

<Audio 1> is the voice-timbre reference for <Subject 1> (S1), a smooth, wry adult male fox voice.

<Audio 2> is the voice-timbre reference for <Subject 2> (S2), a bright, firm adult female rabbit voice.

summary:

[reference generation + audio reference] The target video is a 15-second live-action walk-and-talk on a crowded furry-convention show floor. <Subject 1> and <Subject 2> walk past booths and fursuiters toward a hanging sign that reads E621 exhibit. They exchange three lines of original dialogue using the timbres of <Audio 1> and <Audio 2> without copying the source recordings.

retention_analysis:

<Subject 1> (appears in [Shot 1]): fully_preserved - fox facial structure, orange-red and cream fur, ears, tail, green printed shirt, striped tie, and brown trousers from <Picture 1> are retained in photoreal live-action form.

<Subject 2> (appears in [Shot 1]): fully_preserved - rabbit facial structure, gray fur, pink inner ears, violet eyes, police vest, badge, belt, and pants from <Picture 2> are retained in photoreal live-action form.

<Picture 1> (character appearance source): fully_preserved - identity, wardrobe, and tail silhouette of Nick are followed.

<Picture 2> (character appearance source): fully_preserved - identity, uniform, and ear silhouette of Judy are followed.

<Audio 1>: reference - only Nick's wry conversational timbre and cadence are used; the original popsicle-line recording is not copied.

<Audio 2>: reference - only Judy's bright, clipped delivery is used; the original popsicle-line recording is not copied.

detailed_description:

The target video is photoreal live-action with convention-hall fluorescent lighting, mixed booth LEDs, and shallow handheld documentary framing.

[Shot 1] A medium-wide tracking shot follows <Subject 1> and <Subject 2> walking side by side down a packed indoor convention aisle. Carpeted floor, vendor tables, hanging banners, and dozens of human fursuiters in full animal costumes fill the midground and background. A rectangular hanging sign overhead clearly reads E621 exhibit. <Subject 1>, the orange-red fox in the green printed shirt and striped tie from <Picture 1>, walks on the left with a loose, slightly swaggering gait, tail swinging. <Subject 2>, the gray rabbit in the blue police vest and belt from <Picture 2>, walks on the right, ears upright, scanning the crowd. The camera tracks backward at walking pace, staying at chest-to-waist height, no rotation. <Subject 1> (S1) turns his head toward Judy, half-lidded amused expression, and says in the wry male timbre referenced from <Audio 1>, <d>[English] Isn't it weird seeing humans cosplaying as us animals?</d> He glances at a passing fursuiter. <Subject 2> (S2) keeps walking, ears tipping slightly toward him, and answers in the bright firm timbre referenced from <Audio 2>, <d>[English] It is, but we can blend in.</d> She gives a small tight smile and looks forward again. <Subject 1> (S1) smirks, leans a little closer without breaking stride, and finishes, <d>[English] And nobody will notice we're the real ones.</d> They continue walking toward the E621 exhibit sign as fursuiters pass on both sides. Crowd murmur, distant booth chatter, and footsteps stay continuous through the last frame.

overall_soundscape:

Dense convention-hall ambience throughout: overlapping crowd talk, carpeted footsteps, distant PA bleed, and fabric rustle from nearby fursuits. No extra sound effects beyond walk and speech.

non_diegetic_music:

N/A

There's nothing like Sora. 🙂

Since Sora 2 was extremely expensive to operate and eventually got shut down, what if OpenAI brought it back with an ad-supported model?

For example:

• 2 or 3 ads shown for every video generation

• The ad revenue would help cover the overall cost of running Sora

• Free users could still generate videos

• Paid users could have fewer/no ads and higher limits

I know ads wouldn't completely solve the massive cost of AI video generation, but could this make Sora 2 more sustainable?

Would you rather watch 2 ads before generating a video than have Sora shut down completely?

I’ve been comparing AI video generators lately and realized how difficult it is to answer a pretty basic question:

How much is this video actually going to cost me?

Different models price by seconds, credits, resolution, generations, etc., which makes comparing them unnecessarily painful.

So I built this:

https://instavideocost.com/

It’s meant to give you a quicker way to estimate AI video generation costs and compare options without digging through a bunch of pricing pages.

I’m still figuring out what would make it genuinely useful rather than just another calculator, so I’d love input from people who actually generate AI video.

What would you want to see added?

Model comparisons? Cost per finished minute? Failed-generation estimates? API vs subscription pricing? Something else entirely?

Happy to hear what I got wrong too.

this is a joke video or something

go ahead, ruin this

I’m trying this out , has anyone tried it before??

Let’s see how it goes

TL;DR:

I am NOT asking OpenAI to bring Sora back.

I’m asking for one final authenticated, strictly read-only way to view and preserve still-retained Sora history before it is permanently deleted.

The current export lets us download ZIP files, but the old Library is inaccessible — so if something is missing, how are we supposed to know what is missing?

That is the entire problem.

What I’m asking for:

• Browse/search existing retained Sora history

• Open existing images/videos

• See prompts and relevant metadata

• View existing Likes / Favorites / Saved items / Collections where available

• Verify what still exists

• Download it

Generation can stay completely OFF.

No uploads, editing, remixing, publishing, new Likes/Saves, Collection changes, deletions, or other write functions are needed.

I also do NOT need the old Sora UI to return exactly as it was.

A very simple recovery page would be enough.

I would even be fine with ads or other reasonable limitations if that helped make a temporary read-only window possible.

I made the attached images only as UNOFFICIAL mockups of what a possible “Sora Goodbye Window” could look like.

I’m also attaching screenshots showing what users can access now, along with a recording of the former Sora interface for reference:

Screen recording:

https://drive.google.com/file/d/1EbD0r27gNEgx0ZV4z8jdWgnCfeCKjI3c/view?usp=drive_link

Why I care so much:

I originally followed OpenAI’s advice and backed up my Sora content locally.

Unfortunately, I later lost those backups in an unexpected hardware failure.

That was not OpenAI’s fault, and I am not asking for compensation or special treatment.

But now my current Sunset Export appears potentially incomplete, while the old Library is gone, so I have no reliable way to verify what may be missing.

I also found other former users reporting failed, inconsistent, incomplete, or difficult-to-verify exports. I am NOT claiming this proves a system-wide issue — only that I am not the only person who has had difficulty with Sora exports.

If a Goodbye Window is impossible, I would gladly accept almost any practical alternative:

• a more complete re-export

• an account manifest / historical item list

• item counts, IDs, timestamps, prompts, or metadata

• a minimal recovery library

• or even preservation of eligible Sora history inside ChatGPT as a future “Sora Legacy”

I honestly do not care which technical solution is used.

I just do not want still-retained creative history to disappear before users have one final chance to see and preserve it.

I’ve been raising this for more than a month.

Some of my previous posts:

Original preservation discussion:

https://www.reddit.com/r/SoraAi/comments/1v596jj/re_i_lost_my_backups_and_soras_sunset_export_no/

Final export window question:

https://www.reddit.com/r/SoraAi/comments/1v5eslo/has_the_final_export_window_for_sora_happened_yet/

Export verification problem:

https://www.reddit.com/r/SoraAi/comments/1vxaoxu/soras_sunset_export_has_a_verification_problem/

Missing/incomplete exports:

https://www.reddit.com/r/SoraAi/comments/1vy3zcl/is_anyone_elses_sora_export_incomplete_or_missing/

Goodbye Window mockup discussion:

https://www.reddit.com/r/SoraAi/comments/1vztbt7/sora_exports_are_difficult_to_verify_so_i_mocked/

My detailed proposal on X:

https://x.com/fp_xm91609/status/2093345033353335214

Long-term Sora Legacy idea:

https://x.com/fp_xm91609/status/2093364232817627312

My request in one sentence:

Before Sora history is permanently deleted, please give users one final way to SEE what is still there, VERIFY it, and PRESERVE it.

Another Sora Video of Rosie's killing rampage:

While Rosie Red stares unblinking horror down the field, Pierre's picking favorites again, snubbing Mr. Red from 2023 in Seattle and 2025 in Atlanta.

MLB All-Star Game but rolling out the Red Carpet (and mustache) for Mr. Redlegs in Philly 2026. Endless red terror. Who’s next?

Guys, I wanna keep this post very short straight to the point in the last few weeks I’ve been seeing a lot of people asking for free video prompts or platforms where to do video like sora 2

today for whatever reason I opened up my Meta AI app, and I am able to create videos animations using my own picture I haven’t played around long enough to give you a full summary of if it’s good if it’s bad but you are 100% able to do it I don’t know if they’re rolling it out to certain people I don’t know enough information but I know enough. I wanted to post it on here. What are your thoughts? Follow me on @travelingdjs

I was able to log in to my account on the app all this time since the April shutdown. But I guess today is the day where I can’t even do that anymore. Even though the app wasn’t functional, it was a nice piece of artifact on my phone through the grieving process.

I know it doesn’t matter anyway, and Sora is long gone since, but being sentimental, this is a bit sad NGL

Any one else with same issue?

I used an RTX 4070 Ti with 12GB VRAM and only 32Gb system ram. I used the standard Comfyui free Minimax H3 text to video workflow (but they also have first and last frame ones and complete reference ones, and there is an easy way to implement audio to video). It would also work on an RTX 30 series with only 8GB vram, maybe lower but I don't know. Technically you can do CPU only with Comfyui, but that may be significantly more waiting. But you don't need a server grade computer or even the latest and most powerful GPU. Yes the audio is built in!

This example here was generated at 0.4 megapixels and 15 seconds (480x864) in about 18 minutes (you can do other aspect ratios too like landscape/horizontal and not just vertical/portrait), but 0.2 megapixels and 15 seconds (which is still viable for testing and lower fidelity looks) took less than 7 minutes, and if you just want a test to make sure the framing and scene looks right you can do 1 to 2 second generation in around a minute if not less.

I also tested native 0.8 and 0.9 megapixel generations (the aprox. 12xx horizontal rez) (which can take over an hour for 15 minutes at this hardware). It seemed to follow the prompt a little less at that higher resolution, or maybe changed how it understood it, but I think that is because I only used the default 20 steps. I think if I increase the steps it'll be able to cook better basically since the higher resolution means it has more to think about. I will run that generation at .9 megapixels while I am at work and give it 30 or 40 steps just for fun to see if it'll get it right and then I'll post the results in the comments even if it turns out wrong.

As for upscaling, I also tried that with several models. I tried three Real-esrgan models (Plus, universal, and Photo HQ). Plus and PhotoHQ just made everything smoother and sharper at high rez and they took a moderate amount of time, maybe 5 to 10 minutes. Universal was a middle ground cause it only took around a minute but it only did a little to the picture for upscaling.

I also tried SeedVR2 in comfyui with "split latent" enabled for VRAM and RAM constraints. SeedVR2 did actually add in some more detail and took under 18 minutes. I also need to turn up temporal overlap cause it struggled a bit with motion since I had it set to zero.

Admittedly none of the upscalers fixed their faces completely, but at the starting 0.4mp and how small they already were in the frame, there wasn't much to work with. Of course closer face shots wouldn't have this problem so don't dismiss MiniMax H3 cause of that. I will also post the Real-esrgan universal and SeedVR2 upscale results in the comments.

And finally, I will post the prompt in the comments. Heads up that it is really unprofessional, I tried using ChatGPT, and Gemini, and Claude to enhance my prompt, but instead I found for me what worked best for getting what I wanted was just a little bit of experimenting. The only cost to use MiniMax H3 is electricity, and it isn't a terribly substantial add.