MOSS image 4-Steps workflow… 50 steps w/ blockcache

Combine the depth map control in the same workflow.

Switch it off and disable if you don't need image reference and depth map control.

bypass them if you don't need it, particularly the Depth Anything V2.

Please download the V2 version. V1 version had deleted to avoid any confuse.

The node status of XB-toolbox in V1 version had been expired during comfyui update. and I change it to Dasiwa one for replacement. Both enable switch in Model subgraph and Prompt enhancer fixed and work now.

base and turbo 2 in 1 workflow updated.

combine base and turbo in one workflow.

add switch to the model subgraph, easy switch turbo and base with several switch button.

auto change step and cfg in several switch.

if you're not using DMD turbo lora + base model, (e.g. 4 steps) change the CFG to 3 before run it.

change the prompt enhancer LLM model to Qwen 3.5 4B Q8 to speed up generation time. around 20s in total to enhance prompt + generation.

add photo reference function to the prompt enhancer.

updated user guide in the workflow.

BTW, I had give up the T8 blockcache (but keep the node in subgraph) as the TSR/AuraFlow is fast enough and less image deflect.

if you don't need it, just delete it.

Special thanks:

https://huggingface.co/lodestones/MOSS-Image-Base-8-steps-dmd/tree/main

https://huggingface.co/HauhauCS/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive/tree/main

Add blockcache to speed up 50 steps

Add switch of blockcache and TSR aura flow on the subgraph

Add user guide

### 1. Initial Concept Draft

Input a brief concept into the prompt field, enable the Prompt Enhancer, and run an initial generation using a 4-step low-iteration workflow at 1M resolution.

### 2. Extract Enhanced Prompt

Copy the expanded prompt generated by the Prompt Enhancer and paste it directly into the Positive Prompt field.

### 3. Lock the Desired Seed

Disable the Prompt Enhancer, then continue generating images using the 4-step workflow at 1M resolution until you get a satisfactory result, and note down its Seed number.

### 4. High-Detail Refinement

Input and lock the chosen Seed into a 50-step high-iteration workflow, and re-generate the image at 2M resolution to enhance details and quality.

### 5. Final Upscaling

Use SeedVR2 or Ultimate SD Upscale to upscale the 2M image to your desired final output dimensions.

### 1. 初始概念抽卡

於提示詞欄位輸入簡短的概念描述，開啟提示詞優化器（Prompt Enhancer），並使用 4 步（4 Steps）低步數工作流以 1M 解析度進行首次圖像生成。

### 2. 擷取優化提示詞

複製提示詞優化器自動擴寫出來的詳細提示詞，並將其貼上置換到正向提示詞（Positive Prompt）欄位中。

### 3. 鎖定滿意隨機種子

關閉提示詞優化器，沿用 4 步工作流與 1M 解析度持續進行抽卡，直到生成滿意的圖像為止，並記錄該圖像的隨機種子碼（Seed）。

### 4. 高清細節重製

將滿意圖像的隨機種子碼固定輸入至 50 步（50 Steps）高步數工作流中，並將解析度提升至 2M 重新生成，以補足畫面細節。

### 5. 最終放大輸出

使用 SeedVR2 或 Ultimate SD Upscale 放大工具，將 2M 圖像進一步放大至理想的最終輸出尺寸。

Special Thanks

https://github.com/T8mars/Comfyui-Moss-Image-Blockcache-T8/tree/main

Updated color/depth map control node in generation subgraph and model subgraph.

Updated download link in user guide.

According to the lora creator, lora strength is 0 to 1. 0 means less control, 1 means fully control, could be freely customized by need.

Special thanks to the node creator and lora creator.

https://huggingface.co/darask0/MOSS-Image-color-depth-control/tree/main

https://huggingface.co/depth-anything/Depth-Anything-V2-Base/tree/main

https://github.com/Fannovel16/comfyui_controlnet_aux

A simple MOSS image workflow featuring a locally deployed Prompt Enhancer.

Users can freely choose their preferred LLM deployment. The default model is Qwen3VL-8B Uncensored, which can be easily swapped out for an alternative model based on your specific hardware requirements.

The system defaults to a 4-Step Turbo LoRA, which can also be freely customized or disabled."

MOSS image 的簡單workflow, 配置了本地部署的Prompt Enchancer

可以自由選擇LLM的部署, 預設是Qwen3VL-8B Uncensored, 可自行按硬件需要更換合適的model

預設使用4 Step Turbo Lora, 可自由選擇或關閉。