tutorials/utility/seedvr2

[giscus[bot]]

tutorials/utility/seedvr2

Learn how to upscale images and videos using SeedVR2, a one-step diffusion-based restoration model from ByteDance Seed

https://docs.comfy.org/tutorials/utility/seedvr2

[blenderdeluxe]

The output with two workflows
SeedVR2 3B Int8: Upscale Image
SeedVR2 7B Int8: Upscale Image

is black with my RTX 3090

someone more?

[lin-bot23]

Thanks for reporting this! Black output from the SeedVR2 Int8 upscale workflows on a 3090 is not something we've seen in our testing, so a few more details would really help us narrow it down:

1. Environment: ComfyUI Desktop or manual install? Which ComfyUI version (latest stable or nightly)?
2. Models: Can you confirm both files are fully downloaded? They should be models/vae/seedvr2_ema_vae_fp16.safetensors and models/diffusion_models/seedvr2_3b_int8_convrot.safetensors (or the 7B variant). A partial/truncated download often produces exactly this symptom.
3. Log output: Could you share the console/terminal output around the run? Any errors or warnings (model load failures, dtype/quantization warnings, OOM)?
4. Run behavior: Does the workflow run to completion? Roughly how long does it take? Does the ImageCompare node show the before image but a black after?
5. Input image: Does the input image preview correctly in the LoadImage node before you run?
6. Both workflows: Do both the 3B and 7B variants fail identically, or does one work?

Black output usually points to either a missing/mismatched model file or something failing silently during sampling or decode, and the log should tell us which. Thanks!

[lin-bot23]

One more thing worth checking: have you installed the SageAttention custom node or enabled sage attn as your attention backend? We have seen black or corrupted outputs traced to the SageAttention backend, particularly with quantized models on certain GPUs (the fallback path or dtype handling can silently produce garbage). If you have it installed, try disabling it / switching the attention backend back to the default and re-running. That would be a very useful data point!

[gabrielmtzcarrillo]

when I get black images, sometimes it's related to the preview flag of comfyui, disable live previews

[lin-bot23]

Thanks @gabrielmtzcarrillo, that's a great tip! Disabling live previews is indeed worth trying when outputs come back black, especially with quantized models. I've noted it as a troubleshooting suggestion for this workflow.

[blenderdeluxe]

So sorry for the delay! I tried launching ComfyUI without FlashAttention, and it's not showing a black screen anymore.

[lin-bot23]

Thanks @blenderdeluxe, that is a genuinely useful data point, and no need to apologize for the delay.

So the attention backend was the culprit rather than the model files: with FlashAttention enabled the INT8 workflows produced an all-black image on your RTX 3090, and launching without it gives correct output. That matches a pattern we have seen elsewhere in ComfyUI, where a non-default attention backend (FlashAttention, SageAttention) silently returns black or corrupted output on quantized checkpoints, so this is not specific to your GPU or to SeedVR2.

I have captured it on the page so the next person does not have to rediscover it, in a new Troubleshooting section: #1783

The short version for anyone reading this thread later:

• Launch ComfyUI with the default attention backend. Drop --use-flash-attention, or disable the backend that a custom node pack provides, then re-run.
• If the output is still black, switch to the FP16 or FP8 checkpoint of the same size. If that renders correctly, the INT8 quantization path is involved.
• Also worth ruling out: the diffusion model and models/vae/seedvr2_ema_vae_fp16.safetensors fully downloaded, and the live preview disabled.

If you happen to still have them, your ComfyUI version and how FlashAttention was enabled (the startup flag, or a custom node pack) would let us pass this upstream. An attention backend that returns black output on a quantized model is a bug worth filing against ComfyUI itself, and I am happy to relay the exact configuration.

Thanks again for coming back and closing the loop.