Comfy Console — a simple face for ComfyUI

Prompt

Character / Prompt saves (prompt + LoRAs)

📁 Save folder (blank = flat output)

🔄 ∞ Forever rotates characters

Pick which saved characters to cycle. With ∞ Forever on, each gen loads the next one (prompt + LoRAs). Put wildcards in each character's boxes → endless varied shots of everyone.

⇄ Pull fields from another character ▾

Replace Add to

Character Outfit Pose Location Details Negative LoRAs

Grab only the ticked fields from another saved character into your current prompt — Replace overwrites, Add to appends. Build a character from pieces of several without switching.

🏷️ Tag Shelf ▾ click to insert booru tags

Character → Character box

Outfit → Outfit box

Pose → Pose box

Expression → Pose box

Background → Location box

🎥 Camera / Framing → Pose box (angle + distance)

🎨 Art Style → Details box

💡 Lighting → Details box

🌦 Time & Weather → Location box

🌈 Color / Mood → Details box

Real Danbooru tags (they trigger). Pick one from a list, or hit 🎲 to drop the whole category in as a {wildcard} so every gen rolls a fresh one. Trim options inside the braces to taste. Find more on danbooru.donmai.us; add your own characters with ＋.

Character

Outfit

Pose / Action

Location

Details / Quality

🎲 Wildcards: put {a|b|c} in any box to randomize each gen — e.g. Location = {beach|forest|bedroom|cafe}. Keep the Character box plain (no braces) to lock them. With ∞ Forever, every image rolls fresh options.

Negative

Drop a workflow_api.json here, or click to browse.
In ComfyUI: enable dev mode → Save (API Format).

Drop a generated image here (or click) to read its prompt & settings.
Reads ComfyUI & A1111 PNGs.

Drop an image to fix / upscale (or click)

100% Ctrl+scroll to zoom

🖌 Paint-fix 🧽 Erase 🩹 Clone / Heal 🎨 Remix ⬆ Upscale ✦ Face / Hands ⛶ Extend

Avoid / negative (shared with your main negative)

Drag on the image to paint over the flaw, then Generate. Only the painted area changes.

Brush size

nothing painted yet

♺ Rebuild ✎ Refine · keep shape

Rebuild wipes the area and generates it fresh (turn something into something else). Refine keeps the underlying shape and only repaints the surface (strip hair / a blemish / texture off a feature you want to keep).

What should be there? (optional) Strength Rebuild ~1.0 · Refine ~0.5–0.65 (higher erases the shape, lower keeps the flaw) Soft blend (feathers only the outer seam so the fix melts in — the painted area still changes at full strength)

Reliable removal — no AI, no prompt. Alt-click a clean patch of skin to sample from, then drag over the flaw to paint that clean skin over it. Perfect for a stray navel, mole, seam, or watermark that inpaint won't remove.

Brush size Edge softness higher = smoother blend

alt-click a clean area first

When it looks right, hit Save ▸ — it writes straight to your output folder. No re-render, so nothing else changes.

Turn any dropped image into your look — a photo, someone else's render, or a rough sketch. It runs the image through your checkpoint + LoRAs. Describe the result (or leave it blank to use your Quick prompt), set how much to change, then hit Generate.

Describe the result (blank = your Quick prompt) Change amount low ≈ subtle variation · ~0.5 restyle keeping the pose · 0.75+ mostly ignores the original

Paint over what you want gone — a watermark, a stray object, background clutter. LaMa removes it and fills from the surrounding pixels (no redraw, no hallucination). Cleanest on plain / simple backgrounds.

Edge softness feathers the seam

AI · adds detail Fast · clean

Scale ×

Added detail Tile ControlNet — adds real detail

Big scales eat VRAM — AI mode re-renders in 512px tiles (sharper, slower); Fast is a pure ESRGAN pass (instant, no re-render). With Tile CN on, push Added detail to ~0.35 to actually see the extra detail.

Fix face Fix hands Fix feet Strength

Detects the face / hands / feet and re-renders just those, using your loaded checkpoint + LoRA so it matches your style.

Extend the canvas outward and fill the new space to match — un-crop clipped feet / heads, or change the aspect ratio without squishing.

Extend by (pixels)

Left Right

Top Bottom

What's in the new area? (optional)

Tip: extend one direction at a time for the cleanest result. Uses your checkpoint + LoRA.

Set your checkpoint / LoRA in the Settings panel, then hit Generate ▸. The result shows in the preview and saves to your ComfyUI output folder.

Drop 15–25 images of ONE character (any size — it auto-crops & captions). Give it a one-word trigger, hit Train, wait ~30–60 min. The LoRA lands in your list.

Drop character images here (or click) — you can add more in batches

No images yet.

Trigger word (one word — you'll type this in prompts to summon the character) Epochs (8 = default; it saves every 2 so you can pick the best-baked one)

Good to know Training uses your whole GPU, so you can't generate images while it runs. It captions and buckets automatically — just give it a clean, consistent set (same character, varied poses/outfits/angles). When it's done, the finished LoRA(s) appear in your LoRA list; try the middle epoch (e.g. -000006) at ~0.8 strength with your trigger word.

Settings

Recipe (settings only — no LoRAs)

Checkpoint📋🔄

LoRAs (stackable)🔄

Resolution

🔄 Randomize size each gen (during ∞ Forever)

Steps

CFG

Batch

Sampler

Scheduler

Clip skip

Seed

Detailers — run in order: face → hands → feet

Face

Denoise

Steps

Detail size

Detail size ~512 for distant faces, 768–1024 for close-ups. Lower denoise (~0.35) if busy.

Hands

Denoise

Steps

Fixes mangled fingers (after Face). Denoise 0.35–0.5.

Feet

Denoise

Steps

Fixes mangled toes/feet (after Hands). Denoise 0.35–0.5.

Hi-Res Fix

Upscale ×

Denoise

Steps

Upscale method

Model-upscales & refines (A1111 Hires. fix). Denoise 0.4–0.5. On 8GB keep ≤1.5×.

Ultimate Upscale (4K)

Upscale ×

Denoise

Steps

Tile ControlNet — adds real detail

Tile strength 0.5–0.7; higher = more faithful to source

Tiled 4K, 8GB-safe. Slow — for keepers. Turn off when using ControlNet. Tile-CN lets each tile add coherent detail (raise Denoise to ~0.35 to see more).

ControlNet

Type

Drop a reference image (or click)

Strength

End %

Copies structure from the reference (not the look). Strength 0.7–0.95. One CN model covers all types. Heavy on 8GB — turn off Ultimate/Hi-Res when using it. First use of each type downloads a small preprocessor.

Your images will appear here.

1. Launch ComfyUI (your Desktop shortcut)
2. It auto-connects — the dot turns green
3. Pick a checkpoint, write a prompt, Generate

Log