EchoSingularity/workflows

license: creativeml-openrail-m

WAN 2.2 – I2V Workflow (Optimized for 12GB GPUs)

A fast, clean, and VRAM-efficient Image-to-Video workflow built around WAN 2.2. Fast render times on mid-range GPUs. I tried to keep this simple and easy to use, while maintaining good results. Utilizing well known nodes, and minimizing node bloat. The workflow also has comments everywhere and clear flow.

Ver 1.0 - Base workflow, can do 5 second clips in one iteration. (very fast for 12gb)

Ver 1.1 - More stability, can run 100 times consecutively in 8hrs

Ver 1.2 - Renders 20 second videos. Cleanup of wires.

Ver 1.3 - MMAudio added.

Ver 1.4 - 2x Upscaling, color correction, & sharpening in between passes for quality consistency.

Ver 1.5 - Fixed MMAudio, Updated controls & ability to do 5, 10, 15, & 20 second videos easy. Split RIFE between phases. Fixed prompts. Cleaned up workflow.

---

QWEN Image Edit workflow (Optimized for 12GB GPUs)

Designed to run large AIO QWEN checkpoints (≈28GB) while still generating high-resolution outputs on 12GB VRAM GPUs.

The focus here is:

Image editing / guided edits

Very low step counts

Stable results at low CFG

Aggressive memory management

Clean upscale + post polish

---

Z-Image Turbo workflow (Optimized multi-phase)

Designed to extract maximum detail, edge fidelity, and material realism on 12GB VRAM GPUs. This workflow also includes seed variance to the conditioning so that outputs with the same prompt have more variety similarly to SDXL, Pony, IL models.

This workflow uses controlled sigma shaping, Res-2 samplers, and phased refinement passes to stabilize detail while avoiding common ZIT artifacts like:

Over-etched hair

Shimmering edges

Checkerboard blockiness

CFG-induced harshness

The result is clean, high-contrast outputs that scale well across portraits, fashion, cinematic scenes, and hard-surface material tests.

---

Flux Klein 9B 3 Image Edit

A simple 3-image Klein 9B edit workflow for ComfyUI. This workflow is built for combining multiple image references into one edited output using a prompt-driven FLUX/Klein setup. Drop in three source images, describe how you want them combined, and let the workflow handle the reference conditioning, edit pass, detail refinement, upscaling, and final sharpening.

The setup uses reference latents from all three input images, a two-stage sampling flow, Detail Daemon refinement, Remacri upscaling, resize to 1920px longest side, and CAS sharpening for a cleaner final image.

Good for quick composites, character/object placement, car edits, scene edits, and testing multi-image reference control without building a huge complicated workflow from scratch.

This is a basic V1 workflow, so treat it as a starting point and tweak CFG, denoise, steps, sampler, and prompt strength depending on your images. This workflow uses res2 sampling.

---

Flux.2 Klein 9B Text-to-image High Quality Workflow

This workflow is designed to push FLUX.2 Klein text-to-image generations through a clean 4-stage quality refinement process. It’s built to be easy to follow, easy to tweak, and focused on producing polished final images without turning the graph into a total nightmare.

The workflow starts with a base FLUX.2 Klein generation, then moves into additional sampling refinement using res_2s, Detail Daemon, and a controlled Phase 2 denoise slider. From there it runs FaceDetailer for facial cleanup, then finishes with color balancing, color adjustment, Foolhardy Remacri upscaling, CAS sharpening, and a final resize to 1920px on the longest side.

Step-by-step previews are included for Phase 1, Phase 2, Phase 3, and the final Phase 4 output, along with an image comparer so you can directly compare the first pass against the finished render.

FLUX.2 Klein is already fast and flexible, but this workflow is built to give it a more polished final look with better detail, cleaner faces, sharper output, and a more finished editorial-style render.

---

SDXL 5 Phase/Step Max Detail workflow

This workflow is designed to squeeze everything possible out of SDXL using a 5-pass quality refinement process. It’s easy to use, easy to read, and built for users who want the most polished SDXL output possible.

Each phase adds controlled improvements to detail, clarity, and finish using res_2 samplers, Detail Daemon, FaceDetailer, HandDetailer, upscale, sharpening, and desaturation. Step-by-step preview outputs are included so you can compare every phase and see exactly where the quality gains happen.

SDXL may not be the newest model on the block, but with the right workflow it can still produce beautiful, high-quality images and this setup is built to prove it.

---

Auto IMG Batch Caption workflow automatically generates clean, structured image captions by combining WD14 tagging, Florence-style natural language descriptions, and a custom trigger token for training consistency. The idea behind this workflow is to deliver proven results for easily (one-click) captioning datasets for training. I have made many high quality LORA from the datasets this workflow outputs.

Uses WD14 to extract high-quality tag metadata

Uses Florence to generate a natural-language image description

Injects a custom trigger token at the start of every caption

Outputs both tags + descriptive text in a single caption block

Saves captions to a user-defined folder inside ComfyUI/output

Important Setup Note (VERY IMPORTANT)

You must create a folder inside:

ComfyUI/input/

Example:

ComfyUI/input/Captions

Then select that folder in the caption loader node.

Captions follow this format:

TRIGGER, wd14_tags_here, florence_generated_description_here