import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("CornLogic/10EROS_Int4", torch_dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]Quality - 10Eros_v1.4_DMD_TFO_Int8-Int8mm_Q is 50/50 int8/w4a4-int8mm. 20 second gens ok seems to hold up pretty well.
Balanced -10Eros_v1.4_DMD_TFO_Int8-Int8mm_B is 20/80 int8/w4a4-int8mm. 20 second gens ok, might be better off sticking to 10-15 seconds
Full - 10Eros_v1.4_DMD_TFO_w4a4-Int8mm_F is 0/100 int8/w4a4-int8mm. Does 5 second gens ok, starts to fall apart pretty quickly after that.
All were calibrated on 1920x1024 481 frames 24fps TFO = Transformer Only DMD = DMD lora baked in at 1.0 strength
Made possible by Tsolfils brilliant node https://github.com/tsolful/ComfyUI-INT8-Fast a modification of Bob’s node.
I spent a while going down a rabbit hole splicing int4/int8/bf16 combos in a modified version of comfy-quants, trying to figure out what works before finding Tsolfils node. I’m pleased I was heading in the right direction but his node is probably much better than what I was going to do.
Testing has been unscientific and done in fits and bursts so keep that caveat in mind. It seems any amount of true w4a4 tensors on their own will taint ltx/eros/sulphur. You get swimming textures on any length/size generation. So that's a no go. But packing them in int8mm seems to fix that problem. At first I thought the hard limit was a 50/50 combo of int8 and w4a4 int8 matmul, but once I stopped letting true w4a4 filter into the mix I was able to make a 20/80 mix and 100% w4a4 int8-matmul with no more texture swimming artifacts or vastly diminished at any rate.
Quality - 10Eros_v1.4_DMD_TFO_Int8-Int8mm_Q.safetensor sits below fp8 in my opinion but that's subjective and I have no LPIPS data to back that up.
- Downloads last month
- 39