docs: fix two docstrings that reference names the code does not have

[iridescentWen]

Fixes #14732

What's wrong

Two docstrings name something the function does not have, so following the docs leads nowhere.

File	Line	Docstring said	Reality
models/transformers/transformer_ltx2.py	1449	`peturbation_mask`	perturbation_mask — spelled correctly at :253, :273, :314, :316, :623; the class docstring at :233 also reads "perturbation masking". Only this reference drops the r.
pipelines/deprecated/.../pipeline_stable_diffusion_diffedit.py	885, 888	`source_image`	image — generate_mask has no source_image parameter.

For the second one, the parameter list is:

['image', 'target_prompt', 'target_negative_prompt', 'target_prompt_embeds',
 'target_negative_prompt_embeds', 'source_prompt', 'source_negative_prompt',
 'source_prompt_embeds', 'source_negative_prompt_embeds', ...]

The source_prompt* names those two lines mention alongside it are real, which is what makes the
wrong one misleading rather than obviously broken.

Docstrings only, no behavior change.

Validation

$ uvx ruff check <the 2 changed files>
All checks passed!

$ uvx ruff format --check <the 2 changed files>
2 files already formatted

git diff --stat: 2 files, +3/-3.

Self-review (per CONTRIBUTING)

Ran the .ai/skills/self-review rubric against .ai/review-rules.md:

• Blocking issues: none. Docstring text only.
• # Copied from: checked per changed hunk — forward (transformer_ltx2.py:1366) and
  generate_mask (diffedit:845) carry no # Copied from header, so editing them directly is
  correct rather than fixing an upstream source and running make fix-copies.
  utils/check_copies.py is outside my sparse checkout, so this is the manual equivalent; CI's
  consistency check will confirm.
• Ephemeral context: none added.
• Documentation impact: this is the documentation fix; no docs/ page repeats either name.
• Dead code analysis: N/A.

Deliberately excluded

The same AST pass surfaced two larger clusters that I am not touching, so they read as reviewed
rather than missed:

• alpha_bar, 27 sites across the scheduler files. The docstrings say "Contains a function
  alpha_bar" while the inner function is alpha_bar_fn. That is a conceptual description of what
  the helper does rather than a broken pointer, the sites are linked by # Copied from, and 27 edits
  across that many files is a different kind of change. Happy to send it if you consider it worth
  fixing.
• enhance_t2v / enhance_i2v / generate_enhanced_prompt, 15 sites. These name helpers that do
  not exist anywhere in src/diffusers, so they may be intentional references to external example
  code. I did not want to guess.

The scan needed two rounds of narrowing before it was usable, which is worth stating so the two hits
above are not mistaken for a broad sweep: a module-local symbol table produced far too many hits
because docstrings legitimately reference sibling modules, and lines shaped - `key`: description
enumerate an upstream API's keys rather than the function's own parameters. Both are now excluded.

---

🤖 Written with Claude Code. Both sites were read in context and the
corrected names were taken from the signature and from the surrounding correct spellings.

[stevhliu]

thanks, if you have any more fixes similar to this, feel free to add it to this PR

[iridescentWen]

Thanks — added three more in 68c7626, all the same class (a docstring naming something the
signature does not have), now 5 fixes across 5 files, +6/-6.

The three new ones are parameters that were renamed without the docstring following:

File	Line	Docstring said	Signature
optimization.py	161	num_periods	num_cycles
models/autoencoders/autoencoder_kl_ltx2.py	43	dim	channel_dim
modular_pipelines/z_image/before_denoise.py	100	vae_scale_factor	vae_scale_factor_spatial

Each verified against the real parameter list via AST, and none of the three sits under a
# Copied from header, so editing them directly is correct. ruff check / ruff format --check
clean on all five files.

What I looked at and chose not to include

Since you asked for more of the same, it is worth saying which candidates I rejected and why — three
clusters looked like this class but are not:

• alpha_bar, 27 sites across the schedulers. The docstrings say "Contains a function
  alpha_bar" while the inner helper is alpha_bar_fn. On reading them, these are describing the
  concept the schedule computes, not pointing at a symbol — and 25 of the 27 carry
  # Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar, so the only correct way
  to touch them would be editing scheduling_ddpm.py and running make fix-copies. I left them
  alone; happy to do that separately if you read them as broken pointers rather than prose.
• enhance_t2v / enhance_i2v / generate_enhanced_prompt, 15 sites. These are deliberate
  references to the upstream ltx-core / ltx-pipelines API — the surrounding lines say so
  explicitly. Correct as written.
• tie_last_hidden_states and s_e. The first explains an upstream Qwen3VLTextModel config
  flag; the second is the symbol from equation 12 of the LEDITS++ paper. Both correct.

There is one more genuine cluster I have not put here because it is a different problem class and
would touch 19 files: 27 docstrings write a boolean default as a quoted string, e.g.
force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`) in
pipeline_controlnet_sd_xl.py:222 where the signature is bool = True. Say the word and I will send
it as its own PR, per pipeline family or in one go.