Model Presets

Every time you swap checkpoints you re-tune the same five knobs: sampler, scheduler, steps, CFG, resolution. "Model Presets" exists so you don't have to. Pick a model from a dropdown, pick a prompt vibe, and it hands back a complete set of recommended settings as wires - ready to plug straight into your sampler node. If you keep a stable of checkpoints around, this is the node that makes swapping from FLUX to RealVisXL a two-click affair instead of a session of squinting at model cards.

It's part of the Octo Json Presets pack, and it's the most genuinely useful node in it. The name undersells it a little: it's a lookup table with legs.

How it works

Under the hood the node reads two JSON files bundled in the pack: model_presets.json and prompt_presets.json. Each model entry carries its recommended_settings - sampler, scheduler, steps, CFG, and dimensions - and the node serves the entry for whatever you select. It also maps the model card's internal scheduler label onto ComfyUI's own scheduler names (so a card that says DPMSolverMultistepScheduler comes out as karras, and FLUX's flow-match scheduler maps to beta). That mapping is the part that's actually doing you a favor, because those two naming systems don't line up by default.

You set three things:

• task - image_generation or video_generation, and it flips the model list accordingly. The video side covers HunyuanVideo and SVD.
• model_name - the eight image presets: FLUX.1-dev, FLUX.1-schnell, RealVisXL V4.0, JuggernautXL, DreamShaperXL, RealisticVision V6.0, DreamShaper V8, and Hunyuan-DiT.
• prompt_type - a prompt flavor like photorealistic, portrait, landscape, or couple_at_home.

In return you get scheduler, sampling_method, steps, cfg, width, height, positive_prompt_example, and negative_prompt_example. The sampling knobs wire into your KSampler; the prompts are copy-paste examples with a good default negative built in.

Where people get burned

The presets are community-sensible defaults, not gospel, and a few entries are flat-out wrong for how ComfyUI works. Three things to know:

• dpmpp_2m_karras is not a sampler name. The SDXL presets list it as the sampling_method, but ComfyUI samplers don't carry their scheduler suffix - the sampler is dpmpp_2m and the karras part goes in the scheduler slot, which this node already outputs separately. If you wire sampling_method straight into a KSampler, it won't resolve.
• The FLUX cfg of 3.5 doesn't map to ComfyUI's CFG box. FLUX.1-dev is a flow-matching model; guidance is handled separately, and the standard ComfyUI recipe is CFG 1.0 with guidance ~3.5 on a Guidance node. Plug 3.5 into the plain KSampler CFG and you'll get garbage. This is the classic "presets written by someone whose mental model is SDXL" trap.
• The sampler/scheduler outputs come out wrapped in a list in the current code, which ComfyUI can be picky about feeding into a plain string input. If a connection throws, that's why. Treat sampling_method as a guideline to type in, not a guaranteed wire.

The scheduler mapping itself is sound - the pack sends FLUX to beta, which matches what the community lands on for flow-matching models. And one small JSON quirk: the vintage_photo prompt type has no matching negative in the file, so it silently falls back to the universal negative. Harmless, just know it's happening.

Installing it

Through ComfyUI Manager, search Octo Json Presets and install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/writer-in-fancy-pants/octo_json_presets

Restart and you'll find the node under an octo_presets category. No models to download - this pack only emits settings, so the checkpoints it references (FLUX, Juggernaut, etc.) are things you're expected to have installed already. It's a young, single-commit project, so treat its numbers as a starting point and trust your own eyes over the cards.