Please release an INT4-quantized version of MiniMax-H3
Both diffusion model and text encoder.
It would be good to have a bf16 pruned version, so that users can convert int4
The INT4 is pure garbage. I've tried every combination and they suck. Even the hybrid INT8-INT4. Your best bet is using INT8 for the model and INT4 for the text encoder.
If you're VRAM limited, go with the Q5_K_M GGUF. On 15 passes, it looks pretty good at 768x768.
The INT4 is pure garbage. I've tried every combination and they suck. Even the hybrid INT8-INT4. Your best bet is using INT8 for the model and INT4 for the text encoder.
That’s exactly why I hope the Comfy team can release a well‑optimized INT4 version.
Its possible to convert good working int4 mixed models. I managed to do this with ltx 2.3. its just important to detect/convert the correct layers and not do this randomly. Thats why bf16 (pruned) weights would be interesting for me.
The INT4 is pure garbage. I've tried every combination and they suck. Even the hybrid INT8-INT4. Your best bet is using INT8 for the model and INT4 for the text encoder.
https://huggingface.co/AX1Y2JP/MiniMax-H3-W4A8-ConvRot
This seems to work well, think regular int4 implementation in kitchen isn't updated. Needed the whl and custom node
Int4 at 20 steps and 50 steps. Nothing "garbage" about it. Half the VRAM of int8 for negligible quality difference.
Using https://huggingface.co/tsolful/Minimax_H3_INT4MixedConvRot/blob/main/minimax_h3_fl2va_pruned_INT4BQ.safetensors
This runs natively for me in Comfy, just swapping out Comfy's own int8 model, without any other workflow changes.
But, if someone on the Comfy team were to make an int4, it might surpass this one's quality. (I don't know why so many int4s on HF have been unusable.)
Int4 at 20 steps and 50 steps. Nothing "garbage" about it. Half the VRAM of int8 for negligible quality difference.
Using https://huggingface.co/tsolful/Minimax_H3_INT4MixedConvRot/blob/main/minimax_h3_fl2va_pruned_INT4BQ.safetensors
This runs natively for me in Comfy, just swapping out Comfy's own int8 model, without any other workflow changes.
But, if someone on the Comfy team were to make an int4, it might surpass this one's quality. (I don't know why so many int4s on HF have been unusable.)
i noticed constant shaking in hands when using the int8 model, in your video the hand holding the spoon is shaking also, do you know how many more steps above 20 do i need to stop hands from shaking? thanks.
i noticed constant shaking in hands when using the int8 model, in your video the hand holding the spoon is shaking also, do you know how many more steps above 20 do i need to stop hands from shaking? thanks.
Cannot reproduce. True, the one video included here showed a trembling hand unprompted. Otherwise, I observed no hand shaking regardless of step count, resolution, duration, or prompt.
20 steps, 1088x768, 5 seconds, int4 is perfect for generating videos without hand tremors. (But no different than any other step count.)
20 steps, 1920x1080, 5 seconds, int4 is also perfect for generating videos without hand tremors. (But no different than any other step count.)
Prompt:
5 seconds, 16:9 resolution. A smartphone is held in a woman's left hand, which has smooth skin and long, sparkling emerald fingernails, and with her right hand, she is swiping through the app. The smartphone screen shows an app for browsing magic circles, with the phone time and icons along the top bar, and the home buttons along the bottom bar. On a rustic wooden table in the backround is a boiling cauldron. Each time she swipes the app to display a new magic circle on the screen, a new magical effect erupts from the cauldron, changing the lighting of the room, which is a magic workshop. The smartphone screen begins on a loading animation, and the cauldron is dark, with the room lit by moonlight. She swipes, and on the phone screen appears a red magic circle of arabesques and ornate scrollwork, and fire bursts up from the cauldron. She swipes again, and on the phone screen appears a blue and white magic circle of sharp pointed lines and crystal-like shards, and the cauldron's fire vanishes in a flurry of lightning cracks and showers of electric sparks. She swipes one final time, and the phone screen displays a soft pink magic circle of swirling arcs and bubbling curves, while the lightning vanishes and the cauldron erupts with gently swirling and wafting cherry blossoms. The camera maintains its position for a view of the smartphone and the cauldron, shifting focus from the smartphone, to the cauldron, and then back to the smartphone again. The light source emanates from the cauldron and phone screen, first the warm red of fire, then the harsh, flashing, high-contrast glare of lightning, and finally the magical, pink glow of softly shining cherry blossoms. Cinematic atmospheric documentary, realism, rich texture, dhallow depth of field. The sounds follow the magical effects, with crackling roar of fire, then the cracking boom of thunder, then the delicate flutter of cherry blossoms. End as she sets down the phone and cherry blossoms fill the workshop with a soft light.