## Instructions - Install ComfyUI via portable build or git clone, then install ComfyUI-Manager first for one-click custom nodes. - Install node packs deliberately: ControlNet Aux, Impact Pack, WAS Suite, UltimateSDUpscale — restart after each batch. - Build KSampler chains as Checkpoint > CLIP Text Encode (pos/neg) > KSampler > VAE Decode > Save Image, never skip VAE. - Load LoRAs with LoraLoader stacked after checkpoint: test weight 0.7 first, then sweep 0.5/0.8/1.0 on fixed seed. - Wire ControlNet with preprocessor + Apply node: depth for composition, openpose for pose, canny for line hold. - Version every workflow JSON in git: workflows/v1.2_portrait.json with prompt, seed, model hash in commit message. - Set VRAM flags to match GPU: --normalvram for 8GB, --lowvram for 6GB, --highvram + --preview-method auto for 12GB+. - Use fp16/fp8 checkpoints on low VRAM: fp8 unet + fp16 VAE saves 2GB with negligible quality loss. - Queue in API mode for batches: enable --enable-cors-header, POST prompt JSON, poll history endpoint for PNG. - Batch with efficiency nodes: KSampler Efficient + XY Plot (seed x CFG) overnight before committing to finals. - Snapshot Manager before upgrades: export custom-nodes list, backup models/, test one known workflow after update. - Pairs with ai-video-gen skill: lock character seed there, build node graph here for repeatable renders. ## Examples - Example 1: Portrait chain: SDXL base + Detail LoRA 0.7 + depth ControlNet 0.6, KSampler 30 steps dpmpp_2m, Hires 1.5x. - Example 2: Product batch: 20 prompts queued via API, same seed family 1000-1020, UltimateSDUpscale 2x final. - Example 3: Pose lock: openpose preprocessor on ref photo, pose strength 0.8, prompt delta only for 10 outfits. - Example 4: VRAM rescue: 6GB card switches to --lowvram + fp8 + tiled VAE 512, 768px base then tiled upscale. ## Workflows - Workflow A — New setup: 1. Install portable 2. Manager 3. Node packs 4. Models to checkpoints/loras/controlnet 5. Test workflow. - Workflow B — KSampler final: 1. Draft 20 steps 2. XY sweep 3. Lock seed 4. 40 steps + refiner 5. Hires 6. Save JSON. - Workflow C — LoRA/ControlNet: 1. Load LoRA 0.7 2. Preprocess control image 3. Set strength 4. Test 3 weights 5. Batch. - Workflow D — Versioning: 1. Save API JSON 2. Git commit with seed/model 3. Tag release 4. Export PNG with metadata. ## Templates - VRAM launch flags: ``` # 8GB RTX 3060 Ti daily driver python main.py --normalvram --preview-method auto --use-pytorch-cross-attention # 6GB laptop fallback python main.py --lowvram --fp8-vae --use-split-cross-attention --disable-smart-memory # 24GB workstation batch python main.py --highvram --force-fp16 --enable-cors-header --port 8188 ``` - KSampler API prompt JSON: ```json { "3": {"class_type": "KSampler", "inputs": {"seed": 4242, "steps": 30, "cfg": 6.5, "sampler_name": "dpmpp_2m", "scheduler": "karras", "denoise": 1.0}}, "4": {"class_type": "LoraLoader", "inputs": {"lora_name": "detail_v2.safetensors", "strength_model": 0.7, "strength_clip": 0.7}}, "5": {"class_type": "ControlNetApply", "inputs": {"strength": 0.6, "control_net_name": "depth_sd15.safetensors"}} } ``` - Workflow version header: name portrait-sdxl-v1.3 | model juggernautXL_v9 + fp16 | seed 4242 | git tag wf-1.3. - Manager snapshot: ComfyUI-Manager > Snapshot > Save, commit custom-nodes.txt + workflows/weekly. ## Tools - ComfyUI-Manager — install/update/disable custom nodes and snapshot environments before upgrades. - ControlNet Aux Preprocessors — depth, openpose, canny, lineart in-graph without external preprocessing. - Impact Pack + WAS Suite — face detailer, wildcard prompts, image filters that cut manual editing time. - UltimateSDUpscale — tiled 2x-4x upscale that fits 6GB VRAM where Hires-Fix OOMs. - Civitai + HuggingFace — vetted checkpoints, LoRAs, ControlNet weights with license tags. - ComfyUI API + websocket — queue prompts headlessly for overnight batches and front-end integrations. ## Advanced - Split hires pipeline: base 768px denoise 1.0, then img2img 0.45 denoise at 1.5x to sharpen without recomposing. - Chain two ControlNets: depth 0.6 + openpose 0.8 with preprocessor resolution 512 for stable multi-figure scenes. - Use fp8 unet for Flux: flux1-dev-fp8.safetensors + T5 fp8 clip cuts 24GB requirement down to 12GB workable. - Cache VAE decode with tiled VAE node at 512px tiles with 64 overlap to kill OOM on 4K targets. - Diff workflow JSONs with jq: jq '.nodes | keys' v1.json v2.json to audit changes before sharing publicly. - Route queue across two GPUs: CUDA_VISIBLE_DEVICES=0 port 8188, =1 port 8189, round-robin API POSTs. - Log VRAM per node: enable --verbose, note peak MB per KSampler/Hires combo in workflow README. - Freeze production graphs: duplicate editable graph to _dev, export locked _prod JSON nobody edits by hand. ## Documentation - ComfyUI Wiki — install, CLI flags, Manager usage and custom node troubleshooting index. - ComfyUI-Manager README — snapshot, restore, missing-node install and update-channel policy. - ControlNet Aux docs — preprocessor list, resolution settings, strength tuning per model type. - Civitai model cards — trigger words, recommended CFG/steps, commercial license per checkpoint/LoRA. - KSampler sampler guide — dpmpp_2m vs euler vs heun trade-offs for speed, skin texture, hands.