Spaces:

---

FrameAI4687

/

Omni-Video-Factory

like 872

Running on Zero

Restarted.
Running ComfyUI on ZeroGPU is still experimental. Moving a complex workflow like this to ZeroGPU is a breakthrough, but full stability cannot be guaranteed yet.

no worries, that's life. so it does work now, but the prompt adherence wasn't very good. hopefully u guys can make this as good as the omni image model!
thx for the reply, have a good one

Please add 720p resolution

Please add 720p resolution

I can guarantee that we attempted resolutions of 720p and 1024p, but we encountered crash issues.

The settings we have released for this demo are the highest parameters possible that ensure ZeroGPU will not crash.

Please add 720p resolution

I can guarantee that we attempted resolutions of 720p and 1024p, but we encountered crash issues.

The settings we have released for this demo are the highest parameters possible that ensure ZeroGPU will not crash.

But the output is not 384 or 512p at all.
Something is not right.
Example: i uploaded 768x1152. But the output is 316p or 414p
Ofc you said it from Comfyui so yeah, upscale latent down to side will make it work properly.

I dont know about your setting and the model but to be honest maybe we need to know about sampler, scheduler has been use for this space???
And model shift node, it really important .
In ComfyUI, Model Shift nodes (such as ModelSamplingFlux, ModelSamplingSD3, or ModelSamplingAuraFlow) are used to adjust the sampling distribution of flow-matching models. This "shift" changes how the model behaves during sampling by modifying the timestep/sigma scheduling, which can improve performance, enhance detail, and allow for better control over the generated image or video.
🤔

Please add 720p resolution

I can guarantee that we attempted resolutions of 720p and 1024p, but we encountered crash issues.

The settings we have released for this demo are the highest parameters possible that ensure ZeroGPU will not crash.

But the output is not 384 or 512p at all.
Something is not right.
Example: i uploaded 768x1152. But the output is 316p or 414p
Ofc you said it from Comfyui so yeah, upscale latent down to side will make it work properly.

I dont know about your setting and the model but to be honest maybe we need to know about sampler, scheduler has been use for this space???
And model shift node, it really important .
In ComfyUI, Model Shift nodes (such as ModelSamplingFlux, ModelSamplingSD3, or ModelSamplingAuraFlow) are used to adjust the sampling distribution of flow-matching models. This "shift" changes how the model behaves during sampling by modifying the timestep/sigma scheduling, which can improve performance, enhance detail, and allow for better control over the generated image or video.
🤔

Hey — thanks for the detailed feedback 🙏

Just to clarify how our current setup works:

The output resolution is always reset to multiples of 32 internally (this is required by the model / latent pipeline).
When we mention 384p / 512p, we’re referring to the total pixel budget (e.g. width × height ≈ 384×384 or 512×512), not the exact input resolution.

So if you upload something like 768×1152, it will be rescaled + adjusted to fit within that constraint, which is why the final output may look like ~316p / ~414p on one side.

Regarding your points on sampler / scheduler / model shift — that’s actually very helpful. We’ll take a deeper look into this area and see if there’s room to improve quality and consistency 👍

Appreciate you digging into the details like this!

For me I believe it works well with pixel images with easy story inducing prompts.