Claude ComfyUI workflow-builder skill: 10 node graphs in one prompt

stable-diffusion-image-generation

Companion skill for the prompt-engineering side: tag weighting, CFG tuning, sampler choice.

huggingface

Pull the v1-5-pruned-emaonly.safetensors checkpoint from HF before queueing the workflow.

img2img with denoise control

Same graph as #1 but swap EmptyLatentImage for a LoadImage + VAEEncode chain so KSampler starts from an existing image — denoise=0.55 keeps composition, =0.85 reinvents it.

ForAnyone iterating on a rough sketch, a screenshot, or a previous render they want to nudge.

The prompt

Modify the basic txt2img workflow into img2img. Replace EmptyLatentImage (node 5) with LoadImage reading 'sketch.png' from the input/ folder, plus a VAEEncode node feeding KSampler.latent_image. Set KSampler.denoise to 0.6 (keep composition, repaint surface). Keep the rest. Output as workflow_img2img.json.

What slides.md looks like

{
  "3": { "class_type": "KSampler", "inputs": { "seed": 42, "steps": 20, "cfg": 7.5, "sampler_name": "euler", "scheduler": "normal", "denoise": 0.6, "model": ["4", 0], "positive": ["6", 0], "negative": ["7", 0], "latent_image": ["11", 0] } },
  "10": { "class_type": "LoadImage", "inputs": { "image": "sketch.png" } },
  "11": { "class_type": "VAEEncode", "inputs": { "pixels": ["10", 0], "vae": ["4", 2] } },
  "4":  { "class_type": "CheckpointLoaderSimple", "inputs": { "ckpt_name": "v1-5-pruned-emaonly.safetensors" } },
  "6":  { "class_type": "CLIPTextEncode", "inputs": { "clip": ["4", 1], "text": "oil painting of a stone bridge, 19th century" } }
  /* + nodes 7, 8, 9 unchanged from #1 */
}

One-line tweak

If denoise=0.6 still erases your sketch's lines, drop to 0.4 and bump steps to 30 — fewer surface changes, more refinement.

Pairs with

image-enhancer

Pre-processes the input image (clean noise, lift contrast) so the encoder has a cleaner latent.

nano-banana-pro

When Pro's API rendering quality wins on the editorial shot, swap pipelines without re-prompting.

ControlNet pose with a depth or canny preprocessor

Lock the pose of an input photo into the new render. Adds ControlNetLoader + a preprocessor (DepthAnything or Canny) and a ControlNetApply node between the conditioning and KSampler.

ForDesigners blocking poses for character art, or anyone replacing a stock photo while keeping the silhouette.

The prompt

Build an SD 1.5 ControlNet workflow with depth preprocessing. Use LoadImage('reference.jpg'), DepthAnythingPreprocessor (or MiDaS) → ControlNetApplyAdvanced(strength=0.85) on the positive conditioning. ControlNetLoader = control_v11f1p_sd15_depth.pth. Positive prompt 'a knight in glowing armor, cinematic, volumetric light'. Save as workflow_controlnet_depth.json.

What slides.md looks like

{
  "12": { "class_type": "ControlNetLoader", "inputs": { "control_net_name": "control_v11f1p_sd15_depth.pth" } },
  "13": { "class_type": "LoadImage", "inputs": { "image": "reference.jpg" } },
  "14": { "class_type": "DepthAnythingPreprocessor", "inputs": { "image": ["13", 0], "resolution": 512 } },
  "15": { "class_type": "ControlNetApplyAdvanced", "inputs": { "positive": ["6", 0], "negative": ["7", 0], "control_net": ["12", 0], "image": ["14", 0], "strength": 0.85, "start_percent": 0, "end_percent": 1 } },
  "3":  { "class_type": "KSampler", "inputs": { "positive": ["15", 0], "negative": ["15", 1], "latent_image": ["5", 0], "model": ["4", 0], "seed": 42, "steps": 20, "cfg": 7.5, "sampler_name": "euler", "scheduler": "normal", "denoise": 1 } }
}
# Requires: comfyui_controlnet_aux custom node pack (DepthAnythingPreprocessor)

One-line tweak

Swap DepthAnything for CannyEdgePreprocessor when the input is a line drawing — depth bakes in volume, canny respects strokes.

Pairs with

stable-diffusion-image-generation

Tag-weight and CFG tuning skill that pairs with the conditioning side of the graph.

huggingface

control_v11f1p_sd15_depth.pth lives on lllyasviel/ControlNet-v1-1 — pull via HF.

SDXL base + refiner two-pass workflow

Run the SDXL base model for the first 80% of denoising, then hand the latent to the SDXL refiner for the last 20% — the official two-stage pattern from the SDXL paper.

ForSDXL users who want sharper hands, faces, and small text without leaving ComfyUI.

The prompt

Generate a two-pass SDXL workflow. CheckpointLoaderSimple(sd_xl_base_1.0.safetensors) → CLIPTextEncodeSDXL(positive,negative) → KSamplerAdvanced(steps=25, end_at_step=20, return_with_leftover_noise=enable). Pipe leftover latent into a second KSamplerAdvanced(start_at_step=20) using sd_xl_refiner_1.0.safetensors. 1024x1024, dpmpp_2m_sde / karras. Save as workflow_sdxl_two_pass.json.

What slides.md looks like

{
  "20": { "class_type": "CheckpointLoaderSimple", "inputs": { "ckpt_name": "sd_xl_base_1.0.safetensors" } },
  "21": { "class_type": "CheckpointLoaderSimple", "inputs": { "ckpt_name": "sd_xl_refiner_1.0.safetensors" } },
  "22": { "class_type": "EmptyLatentImage", "inputs": { "width": 1024, "height": 1024, "batch_size": 1 } },
  "23": { "class_type": "KSamplerAdvanced", "inputs": { "add_noise": "enable", "noise_seed": 0, "steps": 25, "cfg": 7.5, "sampler_name": "dpmpp_2m_sde", "scheduler": "karras", "start_at_step": 0, "end_at_step": 20, "return_with_leftover_noise": "enable", "model": ["20", 0], "positive": ["6", 0], "negative": ["7", 0], "latent_image": ["22", 0] } },
  "24": { "class_type": "KSamplerAdvanced", "inputs": { "add_noise": "disable", "noise_seed": 0, "steps": 25, "cfg": 7.5, "sampler_name": "dpmpp_2m_sde", "scheduler": "karras", "start_at_step": 20, "end_at_step": 25, "return_with_leftover_noise": "disable", "model": ["21", 0], "positive": ["6", 0], "negative": ["7", 0], "latent_image": ["23", 0] } }
}

One-line tweak

If hands still drift, drop refiner start_at_step from 20 to 18 — give the refiner a tiny bit more denoising headroom.

Pairs with

ai-image-generation

Higher-level prompt-engineering reference for SDXL aesthetic tags.

huggingface

Both SDXL checkpoints live on stabilityai's HF org.

LoRA stack with three weights tuned

Chain three LoraLoader nodes between the checkpoint and KSampler — a character LoRA at 0.8, a style LoRA at 0.5, a detail LoRA at 0.3 — so each contributes without one swallowing the others.

ForAnyone with a folder of safetensors LoRAs they want to mix without retraining.

The prompt

Build a LoRA stack workflow. After CheckpointLoaderSimple, chain LoraLoader('character_v3.safetensors', strength_model=0.8, strength_clip=0.8) → LoraLoader('cinematic_style.safetensors', 0.5, 0.5) → LoraLoader('skin_detail.safetensors', 0.3, 0.3) → CLIPTextEncode(positive). Use SDXL base, 1024x1024, dpmpp_2m_sde, 30 steps. Output as workflow_lora_stack.json.

What slides.md looks like

{
  "4":  { "class_type": "CheckpointLoaderSimple", "inputs": { "ckpt_name": "sd_xl_base_1.0.safetensors" } },
  "30": { "class_type": "LoraLoader", "inputs": { "lora_name": "character_v3.safetensors", "strength_model": 0.8, "strength_clip": 0.8, "model": ["4", 0], "clip": ["4", 1] } },
  "31": { "class_type": "LoraLoader", "inputs": { "lora_name": "cinematic_style.safetensors", "strength_model": 0.5, "strength_clip": 0.5, "model": ["30", 0], "clip": ["30", 1] } },
  "32": { "class_type": "LoraLoader", "inputs": { "lora_name": "skin_detail.safetensors", "strength_model": 0.3, "strength_clip": 0.3, "model": ["31", 0], "clip": ["31", 1] } },
  "6":  { "class_type": "CLIPTextEncode", "inputs": { "clip": ["32", 1], "text": "<lora:character_v3:0.8> <lora:cinematic_style:0.5> portrait of the character" } }
}

One-line tweak

Sum of strength_model values >1.5 cooks the model — if outputs go melty, scale every weight by 0.7 and re-check.

Pairs with

lora-manager-e2e

Pairs the loader graph with end-to-end LoRA training and management lifecycle.

comfy-cli

comfy CLI installs the LoRA file into models/loras/ before the loader can find it.

Inpainting with a mask brush

Repaint just one region of an image — the SDXL Inpainting checkpoint plus a LoadImage that returns both pixels and a mask. Used for replacing backgrounds, fixing faces, removing objects.

ForPhoto retouchers and product-shot editors who want to change one thing without re-rendering the whole frame.

The prompt

Build an SDXL inpainting workflow. CheckpointLoaderSimple('sd_xl_inpainting_0.1.safetensors'). LoadImage('source.png') returns image+mask (the alpha channel becomes the mask). VAEEncodeForInpaint with grow_mask_by=6. KSampler at denoise=1 inside the masked region only. Positive prompt: 'a small wooden table'. Save as workflow_inpaint.json.

What slides.md looks like

{
  "40": { "class_type": "LoadImage", "inputs": { "image": "source.png" } },
  "41": { "class_type": "CheckpointLoaderSimple", "inputs": { "ckpt_name": "sd_xl_inpainting_0.1.safetensors" } },
  "42": { "class_type": "VAEEncodeForInpaint", "inputs": { "pixels": ["40", 0], "vae": ["41", 2], "mask": ["40", 1], "grow_mask_by": 6 } },
  "3":  { "class_type": "KSampler", "inputs": { "seed": 42, "steps": 25, "cfg": 7.5, "sampler_name": "dpmpp_2m", "scheduler": "karras", "denoise": 1, "model": ["41", 0], "positive": ["6", 0], "negative": ["7", 0], "latent_image": ["42", 0] } }
}
# Mask comes from the alpha channel of source.png (paint it transparent in any editor)

One-line tweak

If the inpainted region has visible seams, raise grow_mask_by from 6 to 12 — bigger feather, smoother blend.

Pairs with

image-editing

Companion skill for masking and selection logic before the inpaint pass.

huggingface

diffusers/stable-diffusion-xl-1.0-inpainting-0.1 is the canonical checkpoint.

Upscale via 4x-UltraSharp + tile pass

Two-stage upscale: ESRGAN-family UpscaleModelLoader (4x_NMKD-Siax_200k or 4x-UltraSharp) for the resolution bump, then a tiled KSampler diffusion pass to add detail without seams.

ForAnyone whose 1024x1024 output needs to land on a 4K poster or a retina hero image.

The prompt

Build a two-stage upscale workflow. UpscaleModelLoader('4x-UltraSharp.pth') → ImageUpscaleWithModel for 4x. Then UltimateSDUpscale (custom node) at tile_size=1024, denoise=0.25, steps=15, dpmpp_2m_sde, with the same SDXL checkpoint and the original prompt for tile-pass conditioning. Save as workflow_upscale_4k.json.

What slides.md looks like

{
  "50": { "class_type": "UpscaleModelLoader", "inputs": { "model_name": "4x-UltraSharp.pth" } },
  "51": { "class_type": "LoadImage", "inputs": { "image": "render_1024.png" } },
  "52": { "class_type": "ImageUpscaleWithModel", "inputs": { "upscale_model": ["50", 0], "image": ["51", 0] } },
  "53": { "class_type": "UltimateSDUpscale", "inputs": { "image": ["52", 0], "model": ["4", 0], "positive": ["6", 0], "negative": ["7", 0], "vae": ["4", 2], "upscale_by": 1, "seed": 42, "steps": 15, "cfg": 7, "sampler_name": "dpmpp_2m_sde", "scheduler": "karras", "denoise": 0.25, "mode_type": "Linear", "tile_width": 1024, "tile_height": 1024, "mask_blur": 8, "tile_padding": 32 } }
}
# Requires: ComfyUI_UltimateSDUpscale custom node

One-line tweak

Drop denoise from 0.25 to 0.18 if tile seams reappear — less per-tile reinvention, more pure upscale.

Pairs with

image-upscaling

Companion skill that wraps the upscale-model picking heuristics (UltraSharp vs Siax vs 4x-Anime).

image-enhancer

Final color/contrast enhancement after the upscale pass.

AnimateDiff motion module — 16-frame loop

Wrap a base SD 1.5 graph with the AnimateDiff motion module to produce a 16-frame, 8-fps looped animation. Adds AnimateDiffLoaderGen1 + a VHS_VideoCombine output node.

ForAnyone shipping animated banners, looping social posts, or sketch animatics.

The prompt

Build an AnimateDiff txt2video workflow. CheckpointLoaderSimple('photonV1.safetensors'). AnimateDiffLoaderGen1(model_name='v3_sd15_mm.ckpt', beta_schedule='sqrt_linear (AnimateDiff)'). EmptyLatentImage with batch_size=16. KSampler steps=20, dpmpp_2m, denoise=1. VHS_VideoCombine at 8fps, output mp4, prefix 'loop'. Save as workflow_animatediff.json.

What slides.md looks like

{
  "60": { "class_type": "AnimateDiffLoaderGen1", "inputs": { "model_name": "v3_sd15_mm.ckpt", "beta_schedule": "sqrt_linear (AnimateDiff)", "model": ["4", 0] } },
  "61": { "class_type": "EmptyLatentImage", "inputs": { "width": 512, "height": 768, "batch_size": 16 } },
  "62": { "class_type": "KSampler", "inputs": { "seed": 42, "steps": 20, "cfg": 7.5, "sampler_name": "dpmpp_2m", "scheduler": "karras", "denoise": 1, "model": ["60", 0], "positive": ["6", 0], "negative": ["7", 0], "latent_image": ["61", 0] } },
  "63": { "class_type": "VHS_VideoCombine", "inputs": { "images": ["8", 0], "frame_rate": 8, "loop_count": 0, "filename_prefix": "loop", "format": "video/h264-mp4", "pingpong": false } }
}
# Requires: ComfyUI-AnimateDiff-Evolved + ComfyUI-VideoHelperSuite

One-line tweak

Set pingpong=true for a back-and-forth loop with no jump cut — perfect for product hero loops.

Pairs with

ai-image-generation

Per-frame prompt strategies that hold motion coherent across the 16 frames.

replicate

When 16 frames at 768x1024 stretch your local VRAM, the same graph runs on Replicate's hosted lane.

IPAdapter face transfer + style transfer combo

Two IPAdapter nodes — one carrying a face reference, one carrying a style reference — both feeding the same KSampler conditioning. Same person, new aesthetic, in one pass.

ForBrand designers, character illustrators, and anyone shipping a consistent persona across many backgrounds.

The prompt

Build an IPAdapter face+style workflow. IPAdapterUnifiedLoader('PLUS (high strength)') with the SDXL base. Two parallel IPAdapter nodes: face image='face_ref.png' weight=0.85 weight_type='style transfer'; style image='style_ref.png' weight=0.6 weight_type='style transfer'. Both pipe into the model side of KSampler. Save as workflow_ipadapter_combo.json.

What slides.md looks like

{
  "70": { "class_type": "IPAdapterUnifiedLoader", "inputs": { "model": ["4", 0], "preset": "PLUS (high strength)" } },
  "71": { "class_type": "LoadImage", "inputs": { "image": "face_ref.png" } },
  "72": { "class_type": "LoadImage", "inputs": { "image": "style_ref.png" } },
  "73": { "class_type": "IPAdapter", "inputs": { "model": ["70", 0], "ipadapter": ["70", 1], "image": ["71", 0], "weight": 0.85, "weight_type": "linear", "start_at": 0, "end_at": 1 } },
  "74": { "class_type": "IPAdapter", "inputs": { "model": ["73", 0], "ipadapter": ["70", 1], "image": ["72", 0], "weight": 0.6, "weight_type": "style transfer", "start_at": 0, "end_at": 1 } },
  "3":  { "class_type": "KSampler", "inputs": { "model": ["74", 0], "positive": ["6", 0], "negative": ["7", 0], "latent_image": ["5", 0], "seed": 42, "steps": 25, "cfg": 6.5, "sampler_name": "dpmpp_2m", "scheduler": "karras", "denoise": 1 } }
}
# Requires: ComfyUI_IPAdapter_plus by cubiq

One-line tweak

If face fidelity slips, raise the face IPAdapter weight to 1.0 and drop style to 0.4 — face is the harder lock.

Pairs with

nano-banana-pro

When the rendered persona needs final-mile typography baked in, hand the IPAdapter output to Pro.

huggingface

h94/IP-Adapter and h94/IP-Adapter-FaceID are both on HF — pull before queueing.

Batch render N variants from a prompt list

Loop one workflow over a list of prompts (or seeds, or LoRA weights) so a single queue produces a labeled grid. Uses the API queue endpoint and a small wrapper script.

ForAnyone running an A/B-style render comparison — palette tests, seed sweeps, prompt ablations.

The prompt

Wrap the basic txt2img workflow in a Python loop that POSTs to /prompt for each entry in PROMPTS = ['a fox in a forest', 'a fox in a city', 'a fox in a desert', 'a fox in snow']. For each, change CLIPTextEncode.text and SaveImage.filename_prefix, post to http://127.0.0.1:8188/prompt, and write the prompt_id to a CSV. Save as run_batch.py.

What slides.md looks like

import json, urllib.request, csv
WORKFLOW = json.load(open('workflow.json'))  # API format
PROMPTS = ['a fox in a forest', 'a fox in a city', 'a fox in a desert', 'a fox in snow']
rows = []
for i, prompt in enumerate(PROMPTS):
    WORKFLOW['6']['inputs']['text'] = prompt
    WORKFLOW['9']['inputs']['filename_prefix'] = f'batch_{i:02d}'
    body = json.dumps({'prompt': WORKFLOW}).encode()
    req  = urllib.request.Request('http://127.0.0.1:8188/prompt', data=body, headers={'Content-Type': 'application/json'})
    pid  = json.loads(urllib.request.urlopen(req).read())['prompt_id']
    rows.append([i, prompt, pid])
csv.writer(open('batch.csv', 'w')).writerows(rows)
# → 4 jobs queued, prompt_ids logged

One-line tweak

Vary `KSampler.seed` instead of `text` to get a four-up seed grid for the same prompt — useful for picking a hero render.

Pairs with

comfyui-request

Pure-API client skill for queueing prompts without opening the GUI.

replicate

Same loop, hosted GPUs — when local VRAM caps your batch size at four.

Community signal

Three voices from people running ComfyUI as their daily pipeline. The first explains why the embedded-PNG-workflow feature has no real equivalent in Automatic1111; the second is the simplest one-line endorsement on launch threads; the third is the everyday ComfyUI-Manager story that resolves the “workflow loaded but seven nodes are red” problem.

“A1111 by nature, has a bunch of disconnected operations in separate tabs and scripts. … Even if the PNG captures all of a generation operation that would be executed by a single launch-button click, its not really equivalent to capturing a whole ComfyUI workflow, which can be the equivalent of a process which would be numerous different tasks in A1111.”

dragonwriter · Hacker News

Comment on the SDXL Turbo + ComfyUI HN thread, explaining why the embedded-PNG-workflow feature has no real equivalent in Automatic1111 — A1111's tabs break the pipeline into manual hand-offs.

“I love that they embed entire workflows into the meta of their images.”

tetris11 · Hacker News

Top-voted reaction on the SDXL Turbo HN thread. The drag-PNG-to-load behavior — every output PNG is a copy of the workflow that produced it — is the single feature most often called out as ComfyUI's killer move.

“Combined with the ComfyUI Manager extensions which provides an index of custom node packages and can install missing ones from a loaded workflow it makes it very easy to get up and running with a new workflow.”

dragonwriter · Hacker News

Companion comment from the same thread. ComfyUI-Manager turns the otherwise-painful 'workflow loaded but seven nodes are red' problem into a one-click install of the missing custom-node packs.

The contrarian take

Not everyone is sold on the node graph as the right surface for day-to-day work. The most precise critique I’ve seen on HN is from LucasPi:

“Parsing ComfyUI workflows is tricky because of the spaghetti node graph.”

LucasPi · Hacker News

From a recent HN comment about parsing ComfyUI workflows.