Changelog
V3.0 [FINAL FOR QWEN IMAGE 2.1]
This will be my final update for the Qwen Image 2.1 lineup. There is little reason to add or tweak further, so I am officially handing the torch over to the community.
The major concept shift in this release is native support for static prompts. I've seen many comments asking: "Why use a PE (Prompt Enhancer) when you can just send a static prompt?" To an extent, that's completely valid - if you're not a writer or artist and simply want to quickly animate your character, static prompts are the way to go.
However, I kept the Auto-PE mode for users who still need granular control over character traits, outfit consistency, and fine details.
PE Model Revamp
A major change in this version is the overhaul of the PE pipeline itself. For many, GGUF proved to be too slow, while running Qwen 3.5 9B natively was too VRAM-heavy. I decided to take a different approach and switch to Qwen 3.5 4B Safetensors.
Why this model?
Tests showed that Qwen 3 VL (4B/8B) struggled with complex instruction following, frequently repeating the system prompt or outputting low-quality sheets.
Switching to Qwen 3.5 4B eliminated the repetition issue, though initial output lacked detail.
I fixed this by streamlining the system prompt. Enabling thinking mode produced noticeably better results: the model reasons through scene composition step-by-step (though it takes a bit more generation time).
Ultimately, prompt generation now takes between 30 seconds and 2 minutes (for int8), with much higher reproducibility and fewer project dependencies.
Two Workflows in One Bundle
I combined V1 and V2 into a single unified package containing two specialized workflows:
..._Production- Use this if you want a polished, highly detailed, and aesthetically rich Character Sheet...._Simple- Use this if you want a streamlined, production-ready Character Sheet built strictly for video animation.
Usage Recommendations
PE Model Selection: You can use other PE models (local or cloud-based). However, for local deployment, I highly recommend Qwen 3.5 4B / 9B depending on your available VRAM. On an 8GB VRAM setup using Qwen 3.5 4B int8, I get around 30 tokens/sec.
Static Prompts: If you don't need manual control over character backstory or precise details, stick to static prompts to speed up generation significantly.
Mode PE Static Prompt Production + - Simple + +
Performance Tip: If PE token generation speed drops unexpectedly, try re-queuing the generation. This appears to be a caching issue, and restarting restores normal speed.
Full Changelog (V3.0)
Dropped GGUF support in favor of native Safetensors.
Rewrote system prompt to optimize instruction adherence for smaller LLM sizes.
Merged V1 (Production) and V2 (Simple) workflows into a single release package.
Resource Links
Qwen 3.5 4B int8: HuggingFace | Qwen 3.5 4B bf16: HuggingFace
Qwen 3.5 9B int8: HuggingFace | Qwen 3.5 9B bf16: HuggingFace
V2.0
In version 2.0, based on community feedback, I redesigned the workflow to be much more task-focused: minimal unnecessary typography, maximum focus on key character elements, body structure, and facial details. This layout makes it drastically easier for video generation models to read character details and maintain consistency.
How It Works
The PE (Prompt Enhancer) generates 2 to 3 main panels:
Facial Expressions
Poses
Equipment & Accessories (optional)
This amount of information is ideal for video models to produce coherent, high-quality results.
Layout Rules & Features:
Faceless Entities: If the object or character does not have a human face (e.g., mask, robot, inanimate object), only a single neutral close-up panel is generated.
Expression Variety: The set of emotional expressions adapts to the character's backstory and description, but a neutral expression always comes first.
Gear & Accessories: If accessories or gear are present, a separate 3rd panel displays them in detail, isolated from the main character.
character_descriptionParameter: Helps the PE better understand character traits and posture. For instance, providing a description like "Ayaka is a slender, fragile girl. She is mostly shy, but occasionally gives a subtle smile" allows the PE to choose matching expressions and poses. This parameter is optional (the PE can infer details strictly from the image), but manually specifying details yields much higher accuracy.
Important Considerations:
Input Image: The workflow performs best when the reference image shows a full-body character facing forward. The PE is intentionally constrained from hallucinating missing body parts unless explicitly specified in the
character_description. This prevents unwanted inconsistencies like altered height, age, or outfit details.Two Workflow Variants: They differ only in how the PE prompt is generated:
For GPUs with <12GB VRAM: Use the standard version (without the
_native_pesuffix).The official
Generate Textnode is currently unoptimized for Qwen 3.5 9B, which can cause prompt generation on lower-end hardware to exceed 30 minutes. The developers are aware of this issue and working on performance fixes.
Summary of Changes (v2.0):
Complete Redesign: Shifted focus from purely stylistic/artistic design sheets to a functional, production-ready Character Design Sheet optimized for video models.
Updated Text Field: Replaced
entity_name(name only) withcharacter_description(full personality & detail specifications).Native PE Support: Added a secondary workflow utilizing the official PE execution method.
Usability: Added clear explanatory comments inside the ComfyUI node graphs.
About this Workflow
This workflow allows you to generate detailed Character Design Sheets from a single reference image. These sheets can later be used as reference frameworks for video generation models like Seedance 2.5 or MiniMax H3.
In my experience, Qwen Image 2.1 is the first open-source image editing model capable of natively generating clean, highly detailed character design sheets right out of the box.
Recommended Settings
CLIP Encoder: Use FP32 / FP16 / BF16 precision for maximum image quality and prompt adherence. Tests show that INT8 and other lower-bit quantizations introduce noticeable artifacts and reduce image clarity. Don't skimp on SSD space — install qwen3vl_8b_bf16.safetensors.
Sampler & Scheduler: Use res_2m + beta to reduce artifacts and improve detail. While I haven't run extensive benchmarks on every combination, this pairing yielded the best results.
Resolution (Megapixels):
3.4 MP — Sweet spot for speed and clarity.
6.0 MP — Superior quality and detail, though generation time increases significantly.
How to Use
Install custom nodes via ComfyUI Manager:
Download Qwen PE I2I or Qwen 3.5 4B/9B weights and place them into
ComfyUI/models/text_encoders/.Run generation: Load your reference image, select the path to the PE model, and start generating.
Useful Resources
Official Qwen Image 2.1 Weight for ComfyUI: HuggingFace
Qwen PE I2I: HuggingFace
Qwen 3.5 4B int8: HuggingFace | Qwen 3.5 4B bf16: HuggingFace
Qwen 3.5 9B int8: HuggingFace | Qwen 3.5 9B bf16: HuggingFace
Description
Dropped GGUF support in favor of native Safetensors.
Rewrote system prompt to optimize instruction adherence for smaller LLM sizes.
Merged V1 (Production) and V2 (Simple) workflows into a single release package.