CodeCow/stable-diffusion-webui-forge
1
0
Fork 0
mirror of https://github.com/lllyasviel/stable-diffusion-webui-forge.git synced 2026-06-04 21:05:48 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

add support for 30b

Browse Source
This commit is contained in:
maybleMyers 2025-12-05 07:58:36 -08:00
parent 4b564f7290
commit 2293a08563
2 changed files with 22 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
backend/sampling/sampling_function.py
View File

@ -438,16 +438,27 @@ def sampling_prepare(unet, x):
is_zimage = getattr(getattr(real_model, 'config', None), 'is_zimage', False)
if is_zimage and hasattr(real_model, 'predictor') and hasattr(real_model.predictor, 'apply_mu_transform'):
# Z-Image uses patch_size=2, so sequence length = (H/2) * (W/2)
image_seq_len = (H // 2) * (W // 2)
real_model.predictor.apply_mu_transform(
seq_len=image_seq_len,
base_seq_len=256,
max_seq_len=4096,
base_shift=0.5,
max_shift=1.15,
)
print(f"Z-Image: Updated mu for {H}x{W} latents (seq_len={image_seq_len}, mu={real_model.predictor.mu:.4f})")
from modules.shared import opts
# Check if user specified a manual shift value
manual_shift = getattr(opts, 'zimage_shift', 0.0)
if manual_shift > 0:
# Use manual shift value directly
real_model.predictor.apply_mu_transform(mu=manual_shift)
print(f"Z-Image: Using manual mu={manual_shift:.4f} for {H}x{W} latents")
else:
# Calculate dynamic shift based on resolution
# Z-Image uses patch_size=2, so sequence length = (H/2) * (W/2)
image_seq_len = (H // 2) * (W // 2)
real_model.predictor.apply_mu_transform(
seq_len=image_seq_len,
base_seq_len=256,
max_seq_len=4096,
base_shift=0.5,
max_shift=1.15,
)
print(f"Z-Image: Auto mu for {H}x{W} latents (seq_len={image_seq_len}, mu={real_model.predictor.mu:.4f})")
# Set up Z-Image CFG handlers
if is_zimage:

1
modules/shared_options.py
View File

@ -201,6 +201,7 @@ options_templates.update(options_section(('sd3', "Stable Diffusion 3", "sd"), {
options_templates.update(options_section(('zimage', "Z-Image", "sd"), {
"zimage_prompt_expansion_max_tokens": OptionInfo(512, "Prompt expansion max tokens", gr.Slider, {"minimum": 128, "maximum": 1024, "step": 64}).info("maximum number of tokens to generate when expanding prompts"),
"zimage_prompt_expansion_temperature": OptionInfo(0.7, "Prompt expansion temperature", gr.Slider, {"minimum": 0.1, "maximum": 1.5, "step": 0.1}).info("higher = more creative, lower = more focused"),
"zimage_shift": OptionInfo(0.0, "Time Shift (mu)", gr.Slider, {"minimum": 0.0, "maximum": 30.0, "step": 0.05}).info("0=auto (resolution-dependent, recommended); manual: 0.5-1.15 typical range; higher=more denoising in early steps"),
"zimage_cfg_normalization": OptionInfo(0.0, "CFG Normalization", gr.Slider, {"minimum": 0.0, "maximum": 3.0, "step": 0.1}).info("rescale CFG result if norm exceeds this multiple of original; 0=disabled, 1.5-2.0 recommended; prevents over-saturation at high CFG"),
"zimage_cfg_truncation": OptionInfo(1.0, "CFG Truncation", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.05}).info("disable CFG after this fraction of steps; 1.0=never truncate, 0.5=disable CFG for last 50% of steps; reduces artifacts"),
}))