New Node: Qwen 2.5 VL Wrapper — QI_QwenVLClipWrapper
Purpose: stabilize VL preprocessing (geometry/resize/color hints) for better color & consistency.
Hook: Official Load CLIP (qwen_image) → Wrapper → your encoder (clip).
Recipes: Portrait center_crop·672·grayscale·0.15·off; Full/vertical letterbox·896·grayscale·0.15·off; Landscape letterbox·896·grayscale·0.15·limit; High‑key letterbox·672·neutral_gray·0.10·off; Color‑change color_mode=original + low desat (≤0.05) or neutralize=false.
Default: letterbox·896·grayscale·0.15·off

QI_TextEncodeQwenImageEdit_Safe (Image Edit Encoder)

Purpose: Encode prompt + input image + VAE into conditioning / image / latent for text–vision mixed editing.
How to use:
- no_resize_pad: letterbox-only (no resample) to preserve pixel consistency.
- pad_mode / grid_multiple: control padding and grid alignment to reduce edge artifacts.
- inject_mode: pixel-reference injection strategy (defaults are fine).
- encode_fp32: enable if VRAM allows for stability.
- prompt_emphasis: strength of prompt adherence (suggest 0.4–0.7).
- vl_max_pixels: CLIP-vision cap (1.4MP) auto-throttled.
Wiring: feed outputs directly into your sampler (e.g., KSampler).

QI_RefEditEncode_Safe (Consistency Edit Encoder)

Purpose: High-consistency editing (portraits, products, layout) with reduced color drift and preserved highlight/shadow detail.
How to use:
- out_width/out_height: lock output size first; the pipeline computes at 32× alignment and crops back, giving pixel-perfect alignment.
- quality_mode: choose among fast / balanced / best / natural for speed/quality trade-off.
- prompt_emphasis: affects only pixel-reference convergence; latent lock remains.
- Built-in Linear BT.709 chroma match + highlight/lowlight attenuation to mitigate bright-scene green shift & highlight clipping; 3MP compute cap for stability.
Wiring: same—pipe conditioning/image/latent to the sampler; latent.qi_pad carries crop metadata for final decoding.

Quick Overview (EN)

No-resize grid padding; CLIP and VAE share the same source image.
reference_pixels come from VAE reconstruction (same domain) for more stable colors; prompt_emphasis is a single knob for text adherence.
Nodes: CN Text Edit, Consistency Edit (RefEdit), VAE Decode.

Quick Wiring: CLIP / IMAGE / VAE → CN Text Edit or Consistency Edit (RefEdit) → (conditioning, image, latent) → sampler → VAE Decode.

Common Parameters:

Quality mode (RefEdit): natural / fast / balanced / best
Injection mode: both / latent only / pixels only (default: both)
Pixel source: VAE reconstruction (default) / original image
Pixel form: color_field_64 (default) / color_field_32 / full pixels
Prompt emphasis: 0–1 (default 0.5; ≥0.8 follows text more)

Nodes

QI_TextEncodeQwenImageEdit_Safe

General text-guided edit encoder. Emphasizes editability while keeping geometry stable via no-resize grid padding and shared CLIP/VAE source. Use prompt emphasis to balance “text adherence ↔ reference stability.”

QI_RefEditEncode_Safe

Two-phase consistency edit encoder. Early (0–0.6) focuses on editability; Late (0.6–1.0) reinforces consistency and detail. A thin high-frequency tail around ~0.985–1.0 cleans edges/hair and suppresses artifacts. Provides quality mode presets (natural|fast|balanced|best) and is Euler / low-CFG friendly by default.

QI_VAEDecodeHQ

High-quality VAE decode. Reads qi_pad metadata and auto-crops back to the original size. No major changes; pair with either encoder.

Usage

Install this package into ComfyUI (place under custom_nodes), no extra dependencies required.
Two sample workflows are provided: basic editing and ControlNet usage.
If you want to use it in your own workflow, just replace TextEncodeQwenImageEdit with this node and use QI_RefEditEncode_Safe in normal cases; for more exploration, use QI_TextEncodeQwenImageEdit_Safe and adjust parameters freely.
Note: Add the sentence “保持人物一致性不变，保持画风光影不变” to your prompt for better results.

Special Thanks

Thanks to 封号 (https://civitai.com/user/Futurlunatic) and PT（娃导 https://github.com/ptmaster ） for rigorous testing; thanks to PT（娃导） for the idea; thanks to 粘土火星（SaturMars：https://github.com/SaturMars ） for technical support; thanks to Dontdrunk （https://github.com/Dontdrunk ），and thanks to the Aiwood 研究院 teammates for their support.