github/ai-toolkit
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Handle unconditional conditioning for ideogram 4 more in line with example code.

Browse Source
This commit is contained in:
Jaret Burkett
2026-06-16 10:23:08 -06:00
parent 86b19589a0
commit f6c0104f25
1 changed files with 13 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
extensions_built_in/diffusion_models/ideogram4/src/pipeline.py
View File

@@ -297,7 +297,11 @@ class Ideogram4Pipeline:
gw = width // (ae_scale * patch) gw = width // (ae_scale * patch)
latent_channels = transformer.config.in_channels latent_channels = transformer.config.in_channels
do_cfg = unconditional_embeds is not None and guidance_scale != 1.0 # Ideogram uses asymmetric CFG: the unconditional branch is image-only
# (no text tokens) with zeroed text features -- it does NOT run a negative
# prompt through the text encoder. So we ignore unconditional_embeds and
# build an empty (0-length) text sequence for the uncond pass below.
do_cfg = guidance_scale > 1.0
if latents is None: if latents is None:
shape = (1, latent_channels, gh, gw) shape = (1, latent_channels, gh, gw)
@@ -311,9 +315,15 @@ class Ideogram4Pipeline:
conditional_embeds.text_embeds, device, dtype conditional_embeds.text_embeds, device, dtype
) )
if do_cfg: if do_cfg:
uncond_feats, uncond_mask = pad_text_features( # Image-only unconditional: zero-length text sequence. predict_velocity
unconditional_embeds.text_embeds, device, dtype # then produces an image-token-only forward pass with zeroed llm
# features, matching the reference's asymmetric CFG.
batch_size = latents.shape[0]
text_dim = cond_feats.shape[-1]
uncond_feats = torch.zeros(
batch_size, 0, text_dim, device=device, dtype=dtype
) )
uncond_mask = torch.zeros(batch_size, 0, dtype=torch.long, device=device)
for t in timesteps: for t in timesteps:
t01 = (t / 1000.0).to(device).expand(latents.shape[0]) t01 = (t / 1000.0).to(device).expand(latents.shape[0])