ComfyUI and Local Image Generation

★ ★ ComfyUI (github.com/comfyanonymous/ComfyUI, GPL-3.0) is the dominant FOSS local image-generation app in 2026. Node-graph workflow; extreme flexibility; supports Stable Diffusion 1.5 / 2 / SDXL / SD3 / SD3.5 / Flux.1 / community fine-tunes. The successor in mindshare to AUTOMATIC1111's webui, which still has a large user base but fell behind on bleeding-edge model support starting around 2024.

For the broader AI section see overview; for video gen see video gen; for hardware see tiers.

ComfyUI

★ ★ ComfyUI — Python + JS; runs locally on Windows / Mac / Linux; web UI on localhost:8188.

• ★ Node graph workflow — every step is a node; chain checkpoints, LoRAs, samplers, ControlNet, IP-Adapter, upscalers.
• ★ First on every new model — SD3, Flux, Hunyuan, Cosmos, all land on ComfyUI before any other UI.
• ★ Workflow sharing — workflows embed in the PNG output; drag a generated PNG into ComfyUI and you have the full pipeline.
• ★ Custom nodes — vast community ecosystem; thousands of extensions.
• ★ API — runs as a server; can be driven programmatically.
• License: GPL-3.0.

Workflow examples worth knowing

• Basic SDXL — checkpoint → CLIP encode → KSampler → VAE decode.
• Flux.1 [dev] — the current quality leader for prompt adherence; needs ~24GB VRAM at full precision (16GB with fp8 / GGUF quants).
• ControlNet pose-to-image — reference image → pose extraction → guided generation.
• IP-Adapter style transfer — reference image → style → new prompt.
• Img2img + inpainting + upscaling chains.
• Animation — AnimateDiff nodes for short videos.

ComfyUI distros

• Vanilla ComfyUI — official.
• ComfyUI Manager — one-click custom node install.
• StableSwarmUI / SwarmUI — friendlier UI on top of ComfyUI.
• Comfy.org distributables — easier install for non-developers.

AUTOMATIC1111 webui

★ AUTOMATIC1111 (github.com/AUTOMATIC1111/stable-diffusion-webui, AGPL-3.0) — the original Stable Diffusion web UI; massive existing user base.

• Mature, polished web UI.
• Form-based rather than node-based; simpler entry for newcomers.
• Slower to add new model support than ComfyUI; SD3 / Flux support arrived later or via forks.
• Active development has slowed since 2024 — still maintained but no longer the bleeding-edge target.

Forge / Stable Diffusion WebUI Forge

• Forge (github.com/lllyasviel/stable-diffusion-webui-forge) — A1111-compatible fork by lllyasviel (creator of ControlNet); bigger optimisations, better Flux support. The active heir to A1111.
• ★ Strong if you want A1111's UX with newer model support.

InvokeAI

★ InvokeAI (github.com/invoke-ai/InvokeAI, Apache 2.0) — alternative; cleaner UI than A1111.

• "Canvas" mode for inpainting / outpainting interactively.
• Friendly for non-technical users.
• Less bleeding-edge than ComfyUI.

Fooocus

★ Fooocus (github.com/lllyasviel/Fooocus, GPL-3.0) — opinionated, simple, "just give me good images."

• One textbox; sensible defaults.
• ★ Best for the "I don't want to learn 10 nodes" user.
• SDXL-flavoured; not the bleeding-edge target.

SD.Next / vladmandic

SD.Next (github.com/vladmandic/automatic, AGPL-3.0) — heavily-modified A1111 fork; broader model support.

Models in 2026

Open base models

• ★ Flux.1 [dev] (Black Forest Labs, 2024) — non-commercial license; current quality leader.
• ★ Flux.1 [schnell] — Apache 2.0; fast 4-step generation; lower quality than [dev].
• ★ Stable Diffusion 3.5 (Stability AI, late 2024) — open-weights; commercial-friendly with Stability license.
• Stable Diffusion XL (SDXL) — older but still standard; many fine-tunes.
• HiDream-I1 — newer 2025 entrant.

Community fine-tunes

• Hugging Face and Civitai (with content caveats) host thousands of community LoRAs and fine-tunes — anime, photorealism, art styles.
• Use Civitai with a strict content filter; not all of it is safe-for-work.

Inpainting / specialized

• OmniGen — multimodal generation with editing.
• OmniControl / ControlNet Plus — composable control.

Hardware reality

Honest tradeoffs vs. paid (Midjourney, ChatGPT-image, DALL·E)

• Flux.1 [dev] is competitive with Midjourney v6/v7 on most subjects; sometimes wins on prompt adherence, sometimes loses on aesthetic.
• Setup pain. ComfyUI requires patience; Midjourney is "type a prompt in Discord."
• Iteration cost. Local is free at the margin; rent-an-H100 if you do it heavily and your machine is too slow.
• Content controls. Local has no central content moderation; this is a feature for legitimate creative work and a vector for abuse — both are real.

Pick this if…

• One pick for serious local image gen: ★ ★ ComfyUI.
• Easier UI than ComfyUI: Forge or InvokeAI.
• Just give me good images, no learning curve: Fooocus.
• You want a simple A1111-flavoured UI: Forge (active) over A1111 (slower-moving).
• Best quality model in 2026: Flux.1 [dev] (license check: non-commercial).
• Best commercial-licensed model: Stable Diffusion 3.5 or Flux.1 [schnell].
• Apple Silicon with MLX: ComfyUI with MLX nodes, or DiffusionBee (closed but free).