Beta 7 Test 2 Native INT8 HQ: Periodic brightness flickering during low-motion sequences
I've noticed periodic brightness flickering with Beta 7 Test 2 Native INT8 HQ (Turbo version), which wasn't present in Beta 6.
Every 17 frames, the entire image dims slightly before returning to normal, sometimes accompanied by faint artifacts. Think a light bulb dimming by 50% then returning to normal right after.
Initially, I thought it was exclusive to 2D anime, but it also occurs in photorealistic footage.
The interesting fact is that it becomes much more noticeable during static or low-motion scenes. My handheld Photorealistic POV generations barely show it, whereas anime scenes with held poses and stationary cameras exhibit it consistently.
In short: Still shots most likely cause the mentioned flicker/dimming, thus it's more apparent in anime-style gens.
I tried the ComfyUI timestep embedding precision fix shown here (https://github.com/Comfy-Org/ComfyUI/issues/15426) (t_emb = self.time_embedder(t_vals)) on model.py, but it didn't help.
I suspect a temporal consistency or VAE boundary issue, possibly related to Beta 7's quantization, though it's just a theory
The checkpoint otherwise produces exceptionally clean results! Distant details are super detailed and sharp even on lower resolutions.
Happens with either Euler or Resmultistep at 8 steps. Using simple Scheduler, 12/3 shift, Comfy Kitchen and Dense attention.
I'll be testing the Test 2 non HQ (Turbo) version next.
Edit: Also happens with the non-HQ version.
I don't believe this is a problem of the quantization format tbh. Some instances of this already popped on the Eros discord, even on the bf16 original source I believe. Seems somehow related to base FastH3 Omni with the Turbos used.
You might want to test some alternatives, like switching to er_sde/sgm_uniform sampler/scheduler pair, or enabling VSA sparsity.
Just as an example or test, you might want to take a look at the workflow I have published for Beta7 on my huggingface: https://huggingface.co/LokkenJP/TenEros_Max_optimized_learned/tree/main/SampleWorkFlowTwoPassOptimized
It's a problem from either running dense or high native resolution. The two pass WF bypasses it.
Got, it! I will test this then. Thanks for the info.