Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -23,6 +23,12 @@ Mixed recipe: 224 block Linears = **140 INT4 + 84 int8_tensorwise fallback**
|
|
| 23 |
`attn.gate`, modulation-class). The shared modulation projector `tproj` and the
|
| 24 |
text-fusion transformer stay bf16.
|
| 25 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
## ⚠️ Runtime requirement
|
| 27 |
|
| 28 |
Needs the `int4_tensorwise` runtime — **not in any released ComfyUI /
|
|
|
|
| 23 |
`attn.gate`, modulation-class). The shared modulation projector `tproj` and the
|
| 24 |
text-fusion transformer stay bf16.
|
| 25 |
|
| 26 |
+
## Samples
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
Same seed/prompt, BF16 top vs INT4-mixed bottom.
|
| 31 |
+
|
| 32 |
## ⚠️ Runtime requirement
|
| 33 |
|
| 34 |
Needs the `int4_tensorwise` runtime — **not in any released ComfyUI /
|