Consider updating the recommendation for HiDream Dev to include a start_percent?

[willhsmit]

I've been using TeaCache with Hidream Dev - it's very useful and the acceleration helps a lot.

Based on my personal experience (in SwarmUI wrapping Comfy), it feels like at the recommended rel_l1_thresh of 1, there's a big difference in output between no TeaCache and TeaCache at threshold 1 with no start_percent, that goes away with a higher start_percent. I wouldn't say the image is broken or distorted, but I wonder if it makes sense to recommend a start_percent on the table in README, for HiDream Dev specifically?

Here's some samples - in each row, the columns are (no teaCache, threshold=1 start_percent=0, threshold=1 start_percent=0.1, threshold=1 start_percent=0.2).

To my eyes,

• At 0 start_percent, there are noticeable structural changes in the image, like the shake or the sauces in the top picture or the outfit of the woman in the hotel, compared to the reference image.
• At 0.1 start_percent, there are fairly subtle differences, like some of the text in the top image or the lines on the floor on the bottom image
• At 0.2 start_percent, these are pretty much gone.

On my 3060 12G, a 25-step render takes 152s with no TeaCache, 83s with start_percent=0, 89s with start_percent=0.1, 93s with start_percent=0.2, so higher start percent does preserve almost all the perf gains of TeaCache.

(edit: this is the city96 Q4_K_S gguf of hidream dev)

[YunjieYu]

Hi @willhsmit. First of all, thank you for your exhaustive experiments on HiDream Dev. Actually, I also tested different start_percent in HiDream Dev, and found that start_percent=0.2 really don't bring any benefits in HiDream Dev bf16 model.

I guess this may be a model-related issue and will not consider updating the start_percent parameter for now. I hope you understand. Finally, thank you again for using ComfyUI-TeaCache!

[willhsmit]

Makes sense, thanks for taking a look.