Comfyui Workflow (Image Generation)
required Custom Nodes:
Ollama Registry: https://registry.ollama.ai/huihui_ai
to run qwen3.5 update the ollama image (0.17.7): docker pull ollama/ollama:latest 2>&1
remove the eclipse-container: docker rm eclipse-ollama 2>&1
if you need help installing nunchaku or how to use the nodes check the readme files in the repo readme folder at github.
llama-cpp-python is needed for the loading_method: GGUF (llama-cpp-python) in the Smart LML Node
nunchaku wheels:
nunchaku model files:
Wildcards
Version 3 uses a new layout and replaces context pipes with KJ's Set/Get nodes plus 2 additional Eclipse nodes (GetFirst and GetAllActive).
Why the layout changed: In the old workflow, changing the model or LoRAs triggered a full re-run including SmartLML's image description, which is slow. The new layout decouples this — SmartLML runs once to describe the image, and that result is reused via Set/Get nodes. Subsequent runs skip the description step when the seed/index is fixed. When shuffling images (-4), SmartLML naturally re-runs since the image changes.
The basic guider is gone — the CFG guider is used for everything (set CFG to 1.0 for Flux, etc.).
Some rgthree nodes have been ported to Eclipse (Fast Bypasser/Muter, Fast Groups Bypasser/Muter, Repeater, Image Compare). They are converted to work with the V3 API and Nodes 2.0, using a different approach — no LiteGraph.registerNodeType interception or subclassing.
GetFirst & GetAllActive: Virtual frontend nodes that extend KJNodes' Set/Get system with priority-based variable resolution.
Get First resolves the first active (not muted/bypassed) SetNode from a prioritized list of variables — ideal for fallback chains (e.g., try LoRA model first, fall back to base model).
Get All Active outputs all active SetNode variables simultaneously, each on its own output slot — perfect for collecting multiple active components.
Both nodes feature:
Type filtering to show only matching SetNode variables
Automatic rename tracking when SetNodes are renamed
Reorder context menu (Move to Top / Up / Down / Bottom, Insert Above)
Green dot indicators showing which variables are currently active
Optional virtual link visualization
Description
Version 3 uses a new layout and replaces context pipes with KJ's Set/Get nodes plus 2 additional Eclipse nodes (GetFirst and GetAllActive).
Requires the latest version of ComfyUI_Eclipse.
add a new node that allows to setup/change the denoise_upscale and the upscale_value that is used in this version of the workflow (latent upscale)
Why the layout changed: In the old workflow, changing the model or LoRAs triggered a full re-run including SmartLML's image description, which is slow. The new layout decouples this — SmartLML runs once to describe the image, and that result is reused via Set/Get nodes. Subsequent runs skip the description step when the seed/index is fixed. When shuffling images (-4), SmartLML naturally re-runs since the image changes.
Since the groups are no longer connected via context, you can arrange them freely.
squeezed the latent upscale functionality into the inital render group.
now it's either Noise Injection or Latent Upscale (both using the same path) otherwise i had to add a 3rd sampler
The basic guider is gone — the CFG guider is used for everything (set CFG to 1.0 for Flux, etc.).
Some rgthree nodes have been ported to Eclipse (Fast Bypasser/Muter, Fast Groups Bypasser/Muter, Repeater, Image Compare). They are converted to work with the V3 API and Nodes 2.0, using a different approach — no LiteGraph.registerNodeType interception or subclassing.
GetFirst & GetAllActive: Virtual frontend nodes that extend KJNodes' Set/Get system with priority-based variable resolution.
Get First resolves the first active (not muted/bypassed) SetNode from a prioritized list of variables — ideal for fallback chains (e.g., try LoRA model first, fall back to base model).
Get All Active outputs all active SetNode variables simultaneously, each on its own output slot — perfect for collecting multiple active components.