`guidance_scale=1.0` no longer disables CFG in LTX-2 pipelines

[rootonchair]

Describe the bug

LTX2Pipeline.do_classifier_free_guidance is:

return (self._guidance_scale > 1.0) or (self._audio_guidance_scale > 1.0)

and since #14447 ("Ltx 2.5", merged 2026-08-11) the __call__ defaults changed:

default	before #14447	after
audio_guidance_scale	None (follows guidance_scale)	7.0
stg_scale	0.0	1.0 (with spatio_temporal_guidance_blocks=[28])
guidance_rescale	0.0	0.7

Consequence: every distilled recipe that passes only guidance_scale=1.0 — including the
LTX-2.3 distilled examples currently in the docs (docs/source/en/api/pipelines/ltx2.md,
"Distilled checkpoint generation" and the condition-pipeline distilled example) and third-party
scripts written before #14447 — now silently runs:

• classifier-free guidance driven by audio_guidance_scale=7.0, with the negative prompt live
  and a doubled transformer batch every step, and
• spatio-temporal guidance (stg_scale=1.0, block 28 perturbed).

The distilled checkpoints are trained for CFG=1 ("runs in 8 steps with CFG = 1"), so this both
doubles compute and degrades output. The 2.5-distilled examples added in #14447 compensate by
passing audio_guidance_scale=1.0 explicitly; the pre-existing 2.3-distilled examples were not
updated and are now broken.

Adding audio_guidance_scale=1.0, stg_scale=0.0, audio_stg_scale=0.0 restores the intended
unguided behavior.

Proposed fix

Since LTX-2 has no use case where one modality is guided while the other is disabled, gate CFG on
both scales, making guidance_scale=1.0 a master off-switch again while the recommended
base-model settings (3.0 video / 7.0 audio) still enable it:

return (self._guidance_scale > 1.0) and (self._audio_guidance_scale > 1.0)

Applied to LTX2Pipeline, LTX2ConditionPipeline, LTX2ImageToVideoPipeline, and the IC-LoRA
variant. No config used anywhere in the repo changes behavior except the broken one: docs use
3.0/7.0 (base — CFG stays on) and 1.0/1.0 (2.5 distilled — off either way); tests use
1.0/1.0; the only 1.0/7.0 occurrences are the stale distilled examples this fixes.

Complementary follow-ups, in whatever combination maintainers prefer:

• Update the LTX-2.3 distilled doc examples to also pass stg_scale=0.0, audio_stg_scale=0.0
  (STG has a separate gate and stays on by default via spatio_temporal_guidance_blocks=[28];
  whether STG-on is intended for distilled checkpoints is worth clarifying — the 2.5-distilled
  examples leave it on).
• Longer-term: gate guidance on a registered is_distilled config flag, as
  Flux2KleinPipeline already does (it warns and skips CFG for distilled checkpoints).

I have the and change ready as a commit and can open a PR if the approach is acceptable.

Affected repos in the wild

Repos matching LTX-2 for diffusers snippets passing guidance_scale=1.0
without audio_guidance_scale; these ship broken examples today:

• diffusers/LTX-2.3-Distilled-Diffusers (official diffusers-org conversion)
• rootonchair/LTX-2.3-Distilled-v1.1-Diffusers
• rootonchair/LTX-2-19b-distilled (also referenced by the docs' distilled example)
• lite-infer/LTX-2.3-Distilled-v1.1-Diffusers-nunchaku-lite-int4-bnb4-text-encoder
• lite-infer/LTX-2.3-Distilled-v1.1-Diffusers-nunchaku-lite-nvfp4-bnb4-text-encoder
• Lightricks/LTX-2

Post-#14447 model cards (Lightricks/LTX-2.5, Lightricks/LTX-2.5-Diffusers) already pass
audio_guidance_scale=1.0 and are unaffected — further evidence the older examples were simply
left behind by the defaults change.

Reproduction

import torch
from diffusers import LTX2Pipeline
from diffusers.pipelines.ltx2.utils import DEFAULT_NEGATIVE_PROMPT, DISTILLED_SIGMA_VALUES

pipe = LTX2Pipeline.from_pretrained("rootonchair/LTX-2.3-Distilled-v1.1-Diffusers", torch_dtype=torch.bfloat16)
pipe.enable_model_cpu_offload()

out = pipe(
    prompt="A flowing river in a forest at golden hour, gentle wind in the leaves.",
    negative_prompt=DEFAULT_NEGATIVE_PROMPT,
    width=768, height=512, num_frames=121, frame_rate=24.0,
    num_inference_steps=8, sigmas=DISTILLED_SIGMA_VALUES,
    guidance_scale=1.0,                      # user believes guidance is off
    generator=torch.Generator("cuda").manual_seed(42),
)
print(pipe.do_classifier_free_guidance)      # True (audio_guidance_scale defaulted to 7.0)
print(pipe.do_spatio_temporal_guidance)      # True (stg_scale defaulted to 1.0)

Logs

System Info

• diffusers: main (reproduced at 0.41.0.dev0)
• torch 2.11.0+cu128, transformers 5.16.1

Who can help?

No response