darksidewalker

/

DaSiWa-WAN2.2-I2V

like 33

Im esp looking forward to BoundBite int8 convrot if you have plans on it xD

It seems that it’s relatively easy to create Int8-ConvRot version using this script:

https://www.reddit.com/r/comfyui/s/iZzcaYTweB

Requires:

• Full precision version of the trained finetune model (not quantized). Public cannot do it as author only published FP8.
• enough RAM to fit the entire model (a lot of RAM)

Thats the thing. At least for me, im gpu poor and i heard that making int8 from fp8 isnt a good idea. Dasiwa-sensei might have the bf16 versions (🤞)

Of course - the FP8 model already has the loss which INT8-ConvRot aims to mitigate. Converting FP8 to INT8-ConvRot just yields an even lossier version. Regardless of the medium, when converting to a lossy format you always want to be using the most lossless input as possible.

We need BF16/FP16 to be able to do it with good quality, otherwise its no use. And i don't think that DaSiWa will share those, because there are guys that take peoples hard work redistribute as something else and don't ever give credit where its due. Also there is StarNodes that makes the whole process very easy, i have already successfully converted my personal int8 convrot stuff and testing int4

thats very true. i asked claude if there was a minimally lossy way to convert FP8 to INT8 and it came up with:

Diff custom FP8 with base Wan FP8 -> Take diff and merge with base FP16 Wan -> convert FP16-ish Wan to INT8..
it wrote me a python file to do all the shenanigans. im honestly surprised how good the gens look. i thought i was going to get complete garbage.

im guessing since most/all wan 'finetunes' are just a bunch of loras merged with base wan2.2 (there maybe some secret sauce from the author too, i dunno), claude figured this way would lose minimal quality.

Im uploading what i made here (https://huggingface.co/Dewprisim/WAN2.2-14B-I2V-INT8) if anyone wants to test it themselves. My laptop is a bit of a potato and hf uploads take forever so it'll be a bit. If DaSiWa or any of the other original authors want me to remove the models i can do that.

YMMV

Are you uploading through git? might be better to use the CLI

the cli. DaSiWa's boundbite v10 should be up now...
Up Next:
wan2.2 remix i2v v3
wan2.2 smoothmix i2v v2

these will probably be the only ones ill do unless the people request more models

thats very true. i asked claude if there was a minimally lossy way to convert FP8 to INT8 and it came up with:

Diff custom FP8 with base Wan FP8 -> Take diff and merge with base FP16 Wan -> convert FP16-ish Wan to INT8..
it wrote me a python file to do all the shenanigans. im honestly surprised how good the gens look. i thought i was going to get complete garbage.

im guessing since most/all wan 'finetunes' are just a bunch of loras merged with base wan2.2 (there maybe some secret sauce from the author too, i dunno), claude figured this way would lose minimal quality.

Im uploading what i made here (https://huggingface.co/Dewprisim/WAN2.2-14B-I2V-INT8) if anyone wants to test it themselves. My laptop is a bit of a potato and hf uploads take forever so it'll be a bit. If DaSiWa or any of the other original authors want me to remove the models i can do that.

YMMV

That's a very interesting thought process there. When I find a chance, I'll test it out.

BTW I see you converted the DaSiWa v10 (Boundbite) but personally I'm a big fan of the latest and greatest v11 (Snatchkiss)

@Dewprisim I can consistently load and run these models without issue:

• SmoothMix V2 I2V High Noise FP8 model
• Wan 2.2 I2V High Noise Int8-ConvRot models (legit quant of the FP16 base model)

But I got OOM many when attempting to load and run your frankenstein'd SmoothMix V2 I2V High Noise Int8-ConvRot.
It seems that to be more hungry for resources. For users with more capable systems, maybe this would actually succeed at what Int8-ConvRot aims to achieve (Faster inference with lesser quality loss), but in my case it just nukes the ability to actually generate at all.

The author uploaded the Int8 convrot versions on Civitai - don't see them here yet

https://civitai.red/models/1981116/dasiwa-wan-22-i2v-14b-or-lightspeed-or-safetensors

@Dewprisim I can consistently load and run these models without issue:

• SmoothMix V2 I2V High Noise FP8 model
• Wan 2.2 I2V High Noise Int8-ConvRot models (legit quant of the FP16 base model)

But I got OOM many when attempting to load and run your frankenstein'd SmoothMix V2 I2V High Noise Int8-ConvRot.
It seems that to be more hungry for resources. For users with more capable systems, maybe this would actually succeed at what Int8-ConvRot aims to achieve (Faster inference with lesser quality loss), but in my case it just nukes the ability to actually generate at all.

Thats odd. Would you happen to have your logs and maybe a screenshot or two of what your vram and ram looked like (I think comfyui has a tool/node for that somewhere) and your comfyui launch flags? Ive been chatting with some people and it looks like at certain higher resolutions, VRAM spikes to high heaven. I use --lowvram (mainly my 6Gb 3060 would die if I didnt) and no issues.

But all seems moot now that dasiwa made their own int8 😃

Thank you for testing my mishmash of model merges. I appreciate it.