Tutorial: Plan — Carry-over triage of the old `/generate` page

Plan — Carry-over triage of the old `/generate` page

Status: decisions made 2026-06-25; no code changed yet. This is the disposition of every control on the legacy classic-server prompt page (src/web/views/generate.pug) — what carries over to the new SPA/core engine, what gets dropped, and in what order. The classic server is frozen and slated for deletion (see the classic-server-read-only directive); this doc is the record of what must be salvaged from it before it goes.

Why the old page felt like a "mash"

It flattens three different concerns into one settings blob, and presets could carry all three at once:

  1. Prompt-text generation / randomization (the engine's job)
  2. What gets sent to the image AI (a provider's job)
  3. Per-image actions ("upscale this file", "make a variation") — not settings at all

Splitting by concern is most of the cleanup. The old hardcoded-defaults → global-overrides → local-overrides (presets) layering is why prompt knobs and image-AI knobs sit side by side.

Disposition

A — Image-AI settings → provider abstraction (deferred, big, later)

Re-homed into the future provider-abstraction project; left as-is until that conversation. Split on a "universal to all providers vs. specific to one" axis.

Control data-path Axis
Target AI settings.mode becomes the provider selector
Stable Diffusion URL imageSettings.url per-provider connection
Generate / Upscale Images settings.generateImages / upscaleImages workflow toggles
Negative Prompt imageSettings.negativePrompt universal-ish (SD/NAI; MJ none)
Images Per Prompt imageSettings.batchCount universal-ish
Seed imageSettings.seed universal
Width / Height imageSettings.width / height universal value, provider-constrained step
Sampler imageSettings.sampler SD-specific
Steps imageSettings.steps SD/NAI
CFG imageSettings.cfg SD/NAI
Restore Faces imageSettings.restoreFaces SD-specific
High-res Detail imageSettings.denoising SD-specific
Variation Percent imageSettings.subseedStrength SD-specific; tied to variation action
All 12 upscale fields upscaleSettings.* SD-WebUI extras/upscaler-specific; likely its own "upscale provider"

B — Emphasis / editing / alternating → rework, provider-dependent (rides on A)

Keep emphasis/de-emphasis as an input concept; the output form is the problem. The old nested-paren (((x))) / [[[x]]] style + booru-era "masterpiece, best quality" framing is dated. Verified 2026-06-25: (word:1.2) weighting still exists in ComfyUI but Flux has minimal sensitivity to weights, and modern models favor natural-language prompting. So:

  • Keep emphasis/de-emphasis as user input intent.
  • Output becomes provider-dependent (each provider emits its own modern form, or none for Flux-like).
  • Cut the global stochastic auto-roll: emphasisChance, emphasisLevelChance, emphasisMaxLevels, deEmphasisChance, keywordEmphasis, keywordEditing + min/max, keywordAlternating + max levels, and chaos (which only scaled that envelope).

C — Replaced by DPL → remove

chaos; keywordCount / keywordMaxCount (→ DPL repeat); autoAddFx (→ v3 wrapper end-block); autoAddArtists / includeArtist / minArtist / maxArtist (→ DPL + wrapper); Anime / Non-Anime Words (→ explicit list selection in DPL).

D — Dead features → remove

All animation: to-animation-file, extend-animation-file, animationStartFrame, animationFrameCount, animationDelay. (Animations never worked well.)

E — Per-image actions → future full image viewer/editor (not built yet)

Image Variation, Re-Roll Image, Re-Roll Which Part, Upscale Image (all data-skip="true" — never persisted to presets anyway). These are "do X to this file" commands; they belong in the image viewer/editor, not the generator.

F — Paths + pipeline → port to the SPA UI/UX as app config

The six folder paths (image / list / expansion / preset / block / prompt-module files) and settings.promptModules (pipeline order). In v3 the pipeline is the engine; surface what's still user-relevant as app config, drop the rest.

Keepers → port into the main SPA UI/UX

  • promptCount (Prompts) — run-level count. Keep.
  • keywordsFilename / artistFilename (Default Keyword/Artist List selectors) — keep, still important.

Park (investigate, don't delete)

  • listEntriesUsedOnce (Unique Entries) + reloadListsOnPromptChange (Each Prompt) — sounds useful; behavior not fully remembered. Investigate before deciding.

Dropped, handled elsewhere

  • promptSalt / promptSaltStart (Auto-add Salt + Starting Number) — bad idea in hindsight; the {salt} token already covers salting. Remove the settings; handle differently if ever needed.

The four sweeps (dependency-ordered — where to begin)

  1. Prune (start here). Drop C, D, and the salt settings. Zero dependencies, removes ~20 controls, makes everything else legible. Lowest risk, highest clarity.
  2. Port the keepers to the SPA UI. promptCount, keyword/artist list selectors, and the F paths as app config. Makes the new generator actually usable.
  3. Park with a paper trail. Record "unique entries — investigate"; tag the E per-image actions for the future image viewer/editor. No building.
  4. Provider abstraction (A) + emphasis rework (B). The big one, later — B's output depends on knowing the provider, so it folds into A. Don't start until ready for that conversation.

See also

  • v3-layers.md — weighted-layer engine; the wrapper that supplies start/end framing (replaces autoAddFx/autoAddArtists).
  • ../reference/prompt-dsl.md — the emphasis/editing/alternating math being reworked, and the {salt} token.
  • next-steps.md — Sweep 1 tracked there as the next concrete item.