Alabo Quality Preset (draft/high)

The habit that wastes the most time in ComfyUI isn't a bad prompt. It's iterating at full resolution - 30 seconds a render - on every idea, when 80% of them are going in the bin anyway. Alabo Quality Preset is one dropdown that makes that impossible to keep doing: flip it to draft, and width, height, and steps all shrink together. Find a keeper, flip to high, and only that winner gets the full render.

It's the sibling node in the same comfyui-alabo-resolution pack as the resolution preset, and like that one it's pure math - no model, no API, no heavy deps. It just emits three numbers that are guaranteed to agree with each other.

How it works

The node holds three orientation presets - portrait 1024x1344, landscape 1344x1024, square 1024x1024 - plus your base_steps. In draft mode it multiplies both dimensions by draft_scale (0.5 by default, so roughly half resolution) and knocks steps down to draft_step_fraction of base_steps (0.4 by default, floored at 8 steps). In high mode you get the full preset exactly. The notes output prints what you actually got, e.g. [draft] portrait (1024x1344) -> 512x672 @ 12 steps, so there's no guessing which resolution a render came from.

The base_steps field is the one to think about: it's meant to be the model's proven full-quality step count, not a guess. The tooltips cite 30 for Z-Image Base and 20 for Flux.2 Klein - which is consistent with how the community actually runs those, where the distilled models want far fewer steps than the base. Set it once per model and the draft recipe scales off something real.

Inputs that matter

• quality - draft or high; the whole point of the node
• orientation - the three fixed aspect presets, 1024-class
• base_steps - your model's real full-quality step count
• draft_scale - draft resolution as a fraction (default 0.5)
• draft_step_fraction - draft steps as a fraction of base_steps

What it wires into

Outputs are width, height, steps, and notes. You'll need to right-click the destination nodes and Convert Widget to Input on their widgets first (drag-off-a-widget also works on newer frontends), then wire width/height into an EmptyLatentImage - or the SD3/Flux2 variants, EmptySD3LatentImage and EmptyFlux2LatentImage, which is the tell that this node is aimed at newer models, not just the SD1.5 world - and steps into KSampler.steps or Flux2Scheduler.steps. Then to scout, set the dropdown to draft, fire off a batch of seeds, and only promote the survivors to high.

This is the same "never iterate at full res" doctrine as the resolution preset, and for stills it's good practice regardless of which model era you're in. Z-Image's own advice is to generate at its ~1024-class native size and multi-pass upscale beyond it; drafting at roughly half that is exactly the cheap-scan step that workflow expects you to do somewhere.

Install

Same pack, same recipe: search "comfyui-alabo-resolution" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/AlaboMPJ/comfyui-alabo-resolution

Then restart. No requirements.txt, so no dependency conflicts to babysit. MIT licensed.

Two gotchas worth knowing: draft steps floor at 8, so on a model that already runs at 8 steps (any of the distilled ones) the draft/high distinction mostly becomes a resolution difference, and that's fine. And if base_steps doesn't match the model you're actually using, the draft recipe is just wrong - the node can't know your model's sweet spot, only you can feed it the right number.