Spaces:

---

SeedOfEvil

/

Pro-Realism-Edit-Studio

like 45

Running on Zero

But I can't see that realistic beauty on his page. So I would like to have it back in This space. Not some psilocybin shroom art

@Matejstein We investigated the history behind the earlier behavior you described and created a separate experimental Space for you to test:

https://huggingface.co/spaces/SeedOfEvil/Pro-Realism-Edit-Studio-Qwen-2511

This is not a cheaper, reduced, or accelerated model intended to save compute. It is actually the opposite: it uses the standard Qwen/Qwen-Image-Edit-2511 transformer in full BF16, without the Rapid-AIO transformer and without FP8 quantization.

Our current hypothesis is that the earlier version may have temporarily been running this standard-Qwen path. That could explain your observation that four steps looked unfinished or vapor-like, while increasing the steps slightly produced a much more photographic result. We cannot confirm that this is exactly the version you remember, so your comparison would be genuinely useful.

To avoid using your Hugging Face ZeroGPU quota too quickly:

1. Sign into Hugging Face before testing.
2. Use one source image and, if possible, the same prompt you used previously.
3. Turn off random seed and keep the same seed for every comparison.
4. Start with 6 steps.
5. If needed, test 8 steps and then 12.
6. Only try the 16-step default if the lower settings do not reproduce the result.
7. Avoid 20 steps initially, run each setting only once, and stop as soon as one matches the photographic look you remember.

Please tell us which step count looked closest. If possible, include the seed, source-image resolution, and resulting image. You do not need to share the prompt if it is private.

The original Pro Realism Edit Studio remains unchanged while we test this hypothesis.

Quick update: we found that the full-BF16 comparison was not actually reaching image generation. ZeroGPU was stalling while trying to bind roughly 57.7 GB of tensors.

I’ve updated the separate test Space to keep the same pinned standard Qwen-Image-Edit-2511 model family—still no Rapid-AIO weights—but its transformer and text encoder now use FP8 so the model can fit on ZeroGPU. This is a compatibility change, not an attempt to substitute a cheaper model or intentionally reduce quality. Its packed footprint is now 29.9 GB, and a real six-step edit completed in approximately 19 seconds.

Updated test Space: https://huggingface.co/spaces/SeedOfEvil/Pro-Realism-Edit-Studio-Qwen-2511

@Matejstein , please try one signed-in test at the six-step default using an image at least 256 pixels on each side. If it works but does not resemble the photographic results you remember, keep the same image, prompt and seed, then try eight steps and finally twelve—once each. Please stop once you find the closest result so you don’t unnecessarily consume your ZeroGPU quota.

If possible, share the output image, step count, seed and source resolution. Your prompt can remain private. FP8 means this cannot prove exact equivalence to the old BF16 behavior, but it should let us properly test whether the standard-Qwen path—not Rapid-AIO—was responsible for the results you preferred.

I've played around with the test space and can't seem to get great results with it. Personally, I've had the best results when manually changing the dimensions to 2048 pixels but that option seems to have been removed. Could you please restore it?

Yea. And Also please give it up to 50 Steps so I can get guidance scale up to heavy edits

@ProtoGhost @Matejstein Thanks — both requests are useful feedback.

We’re reviewing whether 2048-pixel input can be restored safely in the separate test Space. That setting materially increases memory and ZeroGPU cost, so we need to verify it does not bring back the earlier failures before enabling it.

We’ll also evaluate a higher step ceiling as an Advanced/experimental option. We do not want to make 50 steps the default: it will consume substantially more quota, and more steps do not always improve image quality. For now, please compare 6, 8, and 12 steps with the same source image, prompt, and fixed seed. If you can share the source resolution, seed, step count, and output image, that will help us determine whether resolution or sampling depth is the bigger factor. Your prompt can remain private.

No promise yet on the exact limits, but both requests are now part of the test review.

But we are The ones that pay for hugging face subscription, And if is not enough, then we can pay for adding credits. So at least 35 steps would be useful if we want do heavy edits using guidance scale

@ProtoGhost @Matejstein Update: the requested high-cost controls are now available in the separate test Space:

https://huggingface.co/spaces/SeedOfEvil/Pro-Realism-Edit-Studio-Qwen-2511

We restored an experimental 2048×2048 option and added an independent high-step option. Both require explicit opt-in under Advanced Settings; the normal Space defaults remain unchanged.

Our findings so far:

• The first 2048 tests exposed a CUDA memory/allocator failure in the Qwen VAE. We added tiled VAE processing specifically for 2048, and that failure is now mitigated.
• At 2048 and 4 steps, the result looked fuzzy and unfinished.
• At 2048 and 6 steps, the result looked reasonable, but processing took almost a full minute.
• This is consistent with the earlier report that very low step counts can look unfinished and that additional refinement may help. It is promising evidence, but we have not yet proven that higher settings will always produce better images.

To test 2048:

1. Sign into Hugging Face.
2. Open Advanced Settings.
3. Enable experimental high-cost controls.
4. Select 2048 × 2048.
5. Turn off random seed and keep the same image, prompt, and seed.
6. Start at 6 steps. If desired, compare 8, 10, and 12 steps.

The 2048 path is capped at 12 steps because resolution and sampling cost multiply together. For higher-step experiments—up to 50—leave resolution on Auto. Inference steps and true guidance scale are separate controls: increasing steps adds denoising/refinement passes; it does not directly raise guidance strength.

We understand that you are willing to use your own subscription quota or purchased credits for these tests. That choice is available to you, but please expect longer queues, substantially higher quota use, possible timeouts, and diminishing returns. None of these expensive settings are enabled by default.

When reporting results, please include:

• Auto or 2048 resolution
• Step count
• Fixed seed
• Approximate runtime
• Output image
• Any error message

Your prompt can remain private. The original Pro Realism Edit Studio remains unchanged while this separate Space is tested.