feat: Add Ideogram4 model support (CORE-208)

[kijai]

Ideogram 4 support

Adds support for the Ideogram 4 text-to-image model, plus a few generic sampling/conditioning nodes to use it's separate uncondtional model.

Model

• comfy/ldm/ideogram4/model.py — NextDiT/Lumina2-family single-stream DiT. Packs [text, image] into one sequence with block-diagonal segment attention and 3D interleaved MRoPE; runs image-only when no context is given.
• Text encoder: Qwen3-VL-8B with a 13-layer hidden-state tap (53248-dim conditioning).

Nodes (comfy_extras/nodes_ideogram4.py)

• Ideogram4Scheduler — logit-normal sigma schedule.

Generic additions

• DualModelGuider — runs the conditional pass on one model and the unconditional pass on a separate model.
• CFGOverride — override cfg to a fixed value over a [start, end] step range.

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR introduces complete Ideogram 4 model support to ComfyUI. The core implementation adds a single-stream transformer backbone with AdaLN-modulated attention blocks, QKV projections, and RoPE positional encoding. A Qwen3-VL 8B text encoder taps 13 hidden-state layers to produce per-token LLM features injected into text token positions. The model integrates with ComfyUI via detection from state dict keys, registration in supported models, and CLIP type routing. Sampling extensions include a logit-normal sigma scheduler tuned to Ideogram 4's diffusion process and dual-model guidance for separate conditional and unconditional prediction passes. Advanced sampler nodes for CFG override and dynamic guidance scaling complete the feature set.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 3.57% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately summarizes the main change: adding Ideogram4 model support. It is concise, clear, and directly reflects the primary objective of this changeset.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description check	✅ Passed	The PR description clearly relates to the changeset, detailing Ideogram 4 model support with specific components including the transformer backbone, text encoder, and new sampling nodes.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@comfy/ldm/ideogram4/model.py`:
- Line 40: The forward (and any _backbone) signature uses a mutable default
transformer_options={}, which can leak state; change the signature to use
transformer_options=None and inside forward (and _backbone if present) set
transformer_options = {} if None, then use/mutate that local dict. Update
references in comfy/ldm/ideogram4/model.py around the forward method and any
_backbone helper to stop relying on a shared default dict.

In `@comfy/text_encoders/ideogram4.py`:
- Around line 61-65: The flatten currently uses out.permute(0, 2, 3,
1).reshape(b, seq, h * n) which produces hidden-major ordering; to concatenate
each full 4096-dim tap vector in tap-major order change the permutation so taps
are adjacent before the reshape. In encode_token_weights, after getting out,
permute to (0, 2, 1, 3) to produce shape (B, seq, n, h) and then reshape to (b,
seq, n * h) so each tap's full 4096 vector is concatenated in sequence (keep
variables out, b, n, seq, h and returned pooled, extra as before).

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: b6a62744-ae09-4ed5-8de0-ac0c2f54c6f1

📥 Commits

Reviewing files that changed from the base of the PR and between bd7da05 and 8335a17.

📒 Files selected for processing (9)

• comfy/ldm/ideogram4/model.py
• comfy/model_base.py
• comfy/model_detection.py
• comfy/sd.py
• comfy/supported_models.py
• comfy/text_encoders/ideogram4.py
• comfy_extras/nodes_custom_sampler.py
• comfy_extras/nodes_ideogram4.py
• nodes.py

[yamatazen]

Isn't that a closed model?

[MeiYi-dev]

Isn't that a closed model?

Not the latest version. They have it as open weight now.