ComfyUI-Differential-Encode

ComfyUI-Differential-Encode

Seed-sensitive CLIP text encoder for ComfyUI. Designed to fight seed collapse with z-image and similar models that tend to ignore seed changes by injecting controlled, seed-driven perturbations into token embeddings and pooled outputs.

Visual comparison

Same prompt & same seed; only the text encoder differs.

Prompt used: 阳光雀斑女孩：金发扎成松散的侧马尾，脸上有可爱的雀斑，淡妆，笑容灿烂。穿着淡蓝色碎花吊带裙，身材苗条修长。背景是午后的向日葵花田。画风：电影级写实，光影柔和.

<table> <tr> <th align="left">Native CLIP Text Encode</th> <td align="center"><img width="360" alt="Vanilla-1" src="https://github.com/user-attachments/assets/d1acbc83-ae22-4bd1-afc4-d7352dce5f85" /></td> <td align="center"><img width="360" alt="Vanilla-2" src="https://github.com/user-attachments/assets/feb79b9c-697e-4998-be73-984b042451cb" /></td> <td align="center"><img width="360" alt="Vanilla-3" src="https://github.com/user-attachments/assets/0c9faab5-2040-40be-a0dc-daca203965f7" /></td> </tr> <tr> <th align="left">CLIP Text Encode (Differential)</th> <td align="center"><img width="360" alt="Differential-1" src="https://github.com/user-attachments/assets/76186952-98ad-4db9-aeb2-c5212ef275bf" /></td> <td align="center"><img width="360" alt="Differential-2" src="https://github.com/user-attachments/assets/07815b1a-f717-457d-9367-e487350fcf26" /></td> <td align="center"><img width="360" alt="Differential-3" src="https://github.com/user-attachments/assets/d8d82d95-7632-470d-8001-c1609d2db8cd" /></td> </tr> </table>

Workflow example

<img width="1406" height="940" alt="Workflow" src="https://github.com/user-attachments/assets/62ebcac7-6da9-4a84-b7e0-f0e78a042f31" />

Install

Copy or clone this folder into ComfyUI/custom_nodes/ComfyUI-Differential-Encode/ and restart ComfyUI. The node appears as CLIP Text Encode (Differential) under conditioning.

Usage

Replace the standard CLIPTextEncode in your workflow with CLIP Text Encode (Differential). Wire:

• clip: your usual CLIP from the checkpoint loader.
• text: prompt.
• seed: tie to your sampler seed for consistent runs, or vary independently for extra diversity.

Parameters

• noise_std (default 0.18): Gaussian noise per token. Higher = more diversity, too high can drift semantics.
• token_dropout (default 0.10): Probability to drop entire token embeddings.
• per_token_gain_std (default 0.12): Multiplicative jitter per token (1 + N(0,std)) to reshape emphasis.
• global_offset_std (default 0.08): Seeded shift shared by all tokens; nudges conditioning between seeds.
• pooled_noise_std (default 0.05): Noise on pooled_output for models using the pooled branch.
• pooled_global_offset_std (default 0.04): Global shift on pooled_output.
• preserve_magnitude (default True): Re-normalize to original mean/std to avoid blowing up magnitudes.
• orthogonalize_noise (default True): Project noise off the original token directions to keep semantics while adding variety.
• min_std_scale (default 1e-3): Floor for scaling noise so low-variance prompts still get variation.

Tuning tips

• Start with defaults; if images still look identical, raise noise_std to 0.25–0.35 and global_offset_std to ~0.12, or increase token_dropout to 0.15.
• If prompts get too loose, lower per_token_gain_std and keep orthogonalize_noise on.
• For refiner-heavy or pooled-dependent models, bump pooled_noise_std/pooled_global_offset_std together.

Notes

• All perturbations are deterministic per seed.
• Metadata fields differential_* are stored in the conditioning dict for downstream inspection.