comfyui-ideogram-guidance

Ideogram Guidance

Advanced guidance for dual-network diffusion models (e.g. Ideogram 4, which uses a separate conditional and unconditional UNet). One node:

Ideogram DualModelGuider (channelwise + APG) — category Ideogram/guidance, outputs a standard GUIDER (drop-in for ComfyUI's guider socket on SamplerCustomAdvanced).

It combines three techniques to let you push guidance higher for a punchier photoreal look without color burn / oversaturation:

• Channelwise CFG normalization — per-channel std match to the conditional prediction (anti color-burn).
• APG (Adaptive Projected Guidance / orthogonal CFG) — attenuates the part of the guidance vector parallel to the conditional prediction (the saturation driver), keeping the orthogonal part.
• CFG momentum — running average of the guidance direction across steps.

Install

Copy the whole comfyui-ideogram-guidance folder into:

ComfyUI/custom_nodes/comfyui-ideogram-guidance/

Then restart ComfyUI. No extra pip dependencies (uses torch + comfy internals only). Verify on startup that the console lists the node and there are no import errors.

Wiring

cond model   → ModelSamplingAuraFlow(shift=5) → model_cond
uncond model → ModelSamplingAuraFlow(shift=5) → model_uncond   ← SAME shift on BOTH
prompt CLIPTextEncode            → positive
CLIPTextEncode → ConditioningZeroOut → negative
Ideogram DualModelGuider         → SamplerCustomAdvanced (guider input)

Important: both model inputs must go through identical ModelSampling patches. Each network turns the sampler's sigma into its denoised prediction using its own model-sampling; if they differ the two predictions are in mismatched spaces and the combine is incorrect.

cfg is a node input, so you can remove a separate CFGOverride.

Parameters

| Param | Start | Meaning | |-------|-------|---------| | cfg | 4.5 | Guidance scale. Push above vanilla; rescale/APG absorb the saturation. | | channelwise_strength | 0.7 | Per-channel std match. 0=off, 1=full. Anti color-burn. | | apg_eta | 0.0 | Parallel retention. 1.0=plain CFG, 0.0=fully orthogonal (max anti-sat). | | apg_norm_threshold | 0.0 | Optional guidance-norm clamp. 0=off; try 6–12 if highlights clip. | | momentum | 0.0 | Running average across steps. 0=off; APG paper uses ≈ −0.5. |

Identity check: apg_eta=1.0, channelwise_strength=0, momentum=0, apg_norm_threshold=0 reproduces vanilla dual-model CFG at the same cfg. If a fixed-seed A/B matches your stock guider with those settings, the plumbing is correct.

Licensing / provenance

ideogram_dual_guider.py is an independent implementation written from published papers (Lin et al. arXiv:2305.08891 for CFG rescale; Sadat et al. arXiv:2410.02416 for APG + momentum).