Anima, Krea-2, Qwen-Image 2.1, MOSS-Image & YuE2 INT8 Workflows Collection

This repository provides optimized ComfyUI workflows for the Anima, Krea-2, Qwen-Image 2.1, MOSS-Image, and YuE2 model families.

The image-generation workflows cover multiple model architectures and inference approaches. The Anima, Krea-2, Qwen-Image 2.1, and MOSS-Image workflows provide optimized image-generation pipelines with support for quantized model variants.

Anima provides structural control and image-editing capabilities, while Qwen-Image 2.1 supports image generation and editing, mask-based local editing, reference images, and native RGBA output.

MOSS-Image provides text-to-image generation with structured layout control and text rendering. It supports natural-language prompts and structured JSON layouts for controlling the placement of objects and text within generated images.

The Qwen-Image 2.1 workflow also includes optional Spectrum acceleration and native Easy Cache for faster inference.

The repository also includes a YuE2 music-generation workflow for generating complete songs from style prompts and lyrics, including vocals and instrumental accompaniment.

Workflow Files & Overview

The repository contains the following workflow files.

(Note: YYMM.x represents the release year, month, and version number, e.g., 2606.1 or 2607.1.)

#	File Name	Workflow Target	Key Features & Quick Notes
1	Anima_int8-javanoYYMM.x.json	Anima Text-to-Image (INT8)	INT8 image generation with VNCCS, LLLite-based Pose/Canny/Depth control, SAM 3.1 automatic masking, and inpainting.
2	Krea2_int8_t2i-javanoYYMM.x.json	Krea-2 Text-to-Image (INT8)	Quantized image-generation workflow optimized for efficient inference.
3	QwenImage2.1_int8-javanoYYMM.x.json	Qwen-Image 2.1 Text-to-Image / Image Editing (INT8)	Sampling preview, Face Detailer, mask-based editing, reference images, native RGBA output, optional Spectrum acceleration, native Easy Cache, and Ollama-based prompt rewriting.
4	Moss_int8-javanoYYMM.x.json	MOSS-Image Text-to-Image (INT8)	INT8 image generation with structured JSON layout control, spatial composition, and text rendering.
5	YuE2-javanoYYMM.x.json	YuE2 Music Generation	Generates complete songs from style prompts and lyrics using the YuE2-3B model, with vocals and instrumental accompaniment.

Detailed Workflow Breakdown

1. Anima Text-to-Image / Control / Inpainting (Anima INT8)

Designed for high-quality static image generation using the INT8 version of Anima, with additional structural control and automated image editing.

INT8 Inference: Reduced VRAM usage and improved inference efficiency.
Standard Text-to-Image: Generate images directly from text prompts using Anima.
VNCCS Control: Provides Pose Doll-based character and pose control.
LLLite Control: Supports structural guidance using Pose, Canny, and Depth conditioning.
SAM 3.1 Masking: Automatically generates masks for selected subjects or regions.
Inpainting: Combines SAM 3.1 automatic masks with Anima LLLite conditioning for targeted image regeneration.

Anima INT8: https://huggingface.co/Bedovyy/Anima-INT8

Anima LLLite: https://huggingface.co/Comfy-Org/Anima-LLLite

VNCCS: https://github.com/AHEKOT/ComfyUI_VNCCS

SAM 3.1: https://huggingface.co/Comfy-Org/sam3.1

Supported control workflows include:

Standard Text-to-Image
Pose Doll
Canny
Depth
Automatic subject masking
Region-based inpainting

2. Krea-2 Text-to-Image (Krea-2 INT8)

A dedicated INT8 workflow designed to utilize the aesthetic capabilities of Krea-2 within an efficient quantized framework.

INT8 Acceleration: Krea2_int8_t2i-javanoYYMM.x.json leverages INT8 quantization for efficient inference.
Cinematic & Spatial Optimization: Configured to utilize Krea-2's strengths in lighting, realistic rendering, and composition.
Efficient Inference: Quantization reduces memory requirements while maintaining image quality.

3. Qwen-Image 2.1 Text-to-Image / Image Editing (Qwen-Image 2.1 INT8)

A dedicated INT8 ComfyUI workflow for Qwen-Image 2.1, optimized for efficient image generation and image editing.

INT8 Inference: Uses the INT8 version of Qwen-Image 2.1 for reduced VRAM usage and efficient inference.
Sampling Preview: Provides an intermediate sampling preview for monitoring the generation process.
Face Detailer: Adds dedicated face-detail refinement after generation.
Mask Editing: Enables localized image editing by specifying the target region with a mask.
Image Editing: Supports Qwen-Image 2.1 image-editing capabilities.
Reference Images: Supports reference-based image generation and editing.
Native RGBA: Supports native RGBA image generation and transparent backgrounds.
Spectrum Acceleration: Includes optional acceleration using Comfyui-Spectrum-Qwen2.1.
Native Easy Cache: Includes the native ComfyUI Easy Cache node as an alternative inference acceleration method.
Ollama Prompt Optimization: Uses an Ollama-based prompt rewriting workflow optimized around Alibaba's official Qwen-Image prompt-rewriting skills.
Official Prompt Skills: The prompt optimization follows the structure and principles of Qwen-Image's official prompt-rewriting system prompts for Text-to-Image and Image-to-Image editing.

Note: Spectrum and Easy Cache can technically be used simultaneously, but simultaneous use is not recommended. Use either Spectrum or Easy Cache as the acceleration method.

Qwen-Image 2.1: https://huggingface.co/Comfy-Org/Qwen-Image-2.1

Qwen-Image 2.1 Project: https://github.com/QwenLM/Qwen-Image-2.1

Spectrum for Qwen-Image 2.1: https://github.com/awdqwdasdg/Comfyui-Spectrum-Qwen2.1

Supported editing and acceleration workflows include:

Text-to-Image
Image-to-Image
Mask-based local editing
Face Detailer
Reference image editing
Native RGBA / transparent image generation
Spectrum acceleration
Native Easy Cache acceleration

4. MOSS-Image Text-to-Image (MOSS INT8)

A dedicated ComfyUI workflow for MOSS-Image, an image-generation model family developed by the OpenMOSS team.

MOSS-Image is designed for high-quality image generation with particular emphasis on accurate text rendering and structured spatial layout control.

INT8 Inference: Uses an INT8 model variant for reduced memory requirements and efficient inference.
Text-to-Image: Generate images using natural-language descriptions and supported prompt formats.
Structured Layout Control: Use JSON-formatted prompts to describe objects, text elements, and their positions within the image.
Text Rendering: Generate images containing text, including posters, signs, labels, and other typography-focused compositions.
Spatial Composition: Specify object placement and bounding boxes to control the arrangement of visual elements.
Flexible Visual Styles: Supports a range of visual styles, including anime and photorealistic imagery.
ComfyUI Integration: Uses the MOSS-Image integration and its model-loading and sampling workflow.

MOSS-Image-SFT Model: https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT

MOSS-Image Project: https://github.com/OpenMOSS/MOSS-Image

Official ComfyUI Integration Guide: https://github.com/OpenMOSS/MOSS-Image/tree/main/comfyui

Installation

Install the MOSS-Image ComfyUI integration using ComfyUI Manager or follow the official installation instructions.
Download the required INT8 model variant used by this workflow.
Place the model files in the directories specified by the integration documentation.
Load the workflow, enter a prompt, configure the image dimensions and sampling settings, and run generation.

Refer to the official project documentation for the latest installation requirements and model-loading instructions.

JSON Layout Prompting

MOSS-Image supports structured layout prompts for more precise control over composition.

A layout prompt can describe:

The overall scene and visual objective.
Text to render inside the image.
Objects and their visual descriptions.
Bounding boxes and spatial positions for individual elements.

This approach is particularly useful for posters, diagrams, product compositions, title cards, signage, and other images where the relative positioning of text and objects matters.

See the official project documentation for the required JSON schema and coordinate conventions.

Important: The workflow filename identifies this workflow as an INT8 variant. Ensure that the actual model files and loader configuration correspond to the intended quantized model. The MOSS-Image-SFT repository is a model source and should not automatically be assumed to provide the exact INT8 weights required by this workflow.

5. YuE2 Music Generation (YuE2)

A dedicated ComfyUI workflow for YuE2-3B, an open music-generation model from the Multimodal Art Projection (m-a-p) team.

YuE2 generates complete songs from a style prompt and lyrics, producing vocals and instrumental accompaniment.

Text-to-Music: Generate complete songs from style and lyrics.
Vocals & Instrumentation: Produces full songs with vocals and accompaniment.
Symbolic Planning: Uses symbolic musical planning for melody and harmony.
Complete Song Generation: Designed for full-song generation rather than short audio clips.
High-Quality Audio: Supports high-quality stereo music output.

YuE2 Model: https://huggingface.co/Comfy-Org/YuE2

YuE2 Project: https://github.com/multimodal-art-projection/YuE

Typical input structure:

Style: J-pop, female vocal, energetic electronic rock, bright synths, live drums

Lyrics: [Verse] ...

[Chorus] ...

General Notes

These workflows target different model families and are not interchangeable.
Model files, text encoders, VAEs, and custom nodes must match the corresponding model and workflow requirements.
INT8 quantization is used in the designated Anima, Krea-2, Qwen-Image 2.1, and MOSS-Image workflows. It should not be assumed for YuE2.
Custom-node requirements and model-loading procedures may change as the upstream projects evolve.
Refer to each model's official documentation for licensing, installation, and usage details.

Downloads last month: -; Downloads are not tracked for this model. How to track