[Fix] diffusion/flux: use true_cfg_scale as the FLUX CFG weight (#44260)

This commit is contained in:
Wenbo Ji 嵇文博
2026-10-22 21:30:45 +08:00
committed by GitHub
parent 0f7d3a0451
commit afd0d7b47b
4 changed files with 52 additions and 1 deletions
+18
View File
@@ -231,6 +231,24 @@ Use `--lora-path` at startup or the [LoRA management API](/docs/sglang-diffusion
- [`XLabs-AI/flux-furry-lora`](https://huggingface.co/XLabs-AI/flux-furry-lora)
- [`XLabs-AI/flux-RealismLora`](https://huggingface.co/XLabs-AI/flux-RealismLora)
#### 4.2.4 Negative prompts (true CFG)
FLUX.1-dev is guidance-distilled: its guidance embedding comes from the server's `--embedded-cfg-scale` (3.5 for FLUX.1), and the requests above run one DiT forward per step. A negative prompt turns on classifier-free guidance, which adds an unconditional forward per step:
- `--true-cfg-scale` sets the CFG weight, as in diffusers `FluxPipeline`.
- Without `--true-cfg-scale`, the negative prompt uses `--guidance-scale` as the CFG weight.
```bash
sglang generate \
--model-path black-forest-labs/FLUX.1-dev \
--prompt "A red bicycle leaning against a stone wall beside a quiet lake at sunset" \
--negative-prompt "blurry, distorted, low quality" \
--true-cfg-scale 2.0 \
--width 1024 --height 1024 --num-inference-steps 50 --seed 42
```
Keep the weight moderate. On top of the distilled guidance, CFG over-exposes FLUX.1-dev in diffusers as well as in SGLang: at 1024x1024 and 28 steps, about 45% of pixels clip to 0 or 255 at a weight of 4, and 8-15% at 1.5-2.
## 5. Benchmark
### 5.1 Speedup Benchmark
@@ -125,6 +125,12 @@ class FluxPipelineConfig(ImagePipelineConfig):
def prepare_sigmas(self, sigmas, num_inference_steps):
return self._prepare_sigmas(sigmas, num_inference_steps)
def get_classifier_free_guidance_scale(self, batch, guidance_scale: float) -> float:
# true_cfg_scale is the CFG weight when given, as in diffusers FluxPipeline.
if batch.true_cfg_scale is not None:
return batch.true_cfg_scale
return guidance_scale
def prepare_latent_shape(self, batch, batch_size, num_frames):
height = 2 * (
batch.height // (self.vae_config.arch_config.vae_scale_factor * 2)
@@ -1411,7 +1411,7 @@ class SamplingParams:
"--true-cfg-scale",
type=float,
dest="true_cfg_scale",
help="True CFG scale for models that distinguish distilled guidance from standard CFG (e.g., Qwen-Image)",
help="True CFG scale for models that distinguish distilled guidance from standard CFG (e.g., Qwen-Image, FLUX.1)",
)
add_argument(
"--guidance-rescale",
@@ -0,0 +1,27 @@
import unittest
from types import SimpleNamespace
from sglang.multimodal_gen.configs.pipeline_configs.flux import (
Flux2PipelineConfig,
FluxPipelineConfig,
)
class TestFluxTrueCFGScale(unittest.TestCase):
def test_true_cfg_scale_is_the_cfg_weight(self):
for config in (FluxPipelineConfig(), Flux2PipelineConfig()):
with self.subTest(config=type(config).__name__):
batch = SimpleNamespace(true_cfg_scale=4.0)
self.assertEqual(
config.get_classifier_free_guidance_scale(batch, 3.5), 4.0
)
def test_guidance_scale_is_the_cfg_weight_without_true_cfg_scale(self):
batch = SimpleNamespace(true_cfg_scale=None)
self.assertEqual(
FluxPipelineConfig().get_classifier_free_guidance_scale(batch, 3.5), 3.5
)
if __name__ == "__main__":
unittest.main()