Nodes/Eric_Krea2/Eric Krea2 Sweep Plan
ComfyUI Node

Eric Krea2 Sweep Plan

One queue item, a whole grid of samplers and schedules

By EricRollei·Created 4 months ago·Updated 2 days ago· 23
Eric Krea2 Sweep Plan
    • sweep
    ◄modegrid►
    ◄stage_scopes1►
    ◄samplers►
    ◄schedulers►
    ◄pair_lockfalse►
    ◄steps►
    ◄windows►
    ◄etas►
    ◄noise_types►
    ◄cfgs►
    ◄shift_modes►
    ◄mus►
    ◄combos_file►
    ◄combos_json►
    ◄seed_policyfixed►
    ◄seeds►
    ◄quick_screenfalse►
    ◄screen_megapixels1.00►
    ◄score_metricstrue►
    ◄sort_byplan_order►
    ◄tile_px448►
    ◄max_combos0►
    ◄output_folder►
    ◄hybrid_steps►
    ◄megapixels►
    ◄lora_modenone►
    ◄lora_files►
    ◄lora_strengthscard►
    ◄lora_target1►
    ◄include_baselinetrue►
    ◄keep_panel_stackfalse►
    ◄prompts►
    ◄sigma_profiles►
    ◄lora_triggerssaved pick►
    ◄enabledtrue►
    ◄sweep_presetcustom►
    ◄preset_notes►
    ◄run_notes►
    ◄share_sheetsstrips►
    ◄share_backgroundblack►
    ◄stack_partseach LoRA alone►
    ◄sde_bands►
    ◄churn_forms►
    ◄hybrid_handoffs►
    ◄trigger_joins►

    Every fine-tune has a preferred sampler, scheduler, step count and mu - and all of them are really proxies for one thing: how far that model's training eroded the turbo distillation. Instead of guessing, this node measures. Wire Eric Krea2 Sweep Plan into Ultra V2's sweep input (the last socket), fill in the axes you want to test, and one queue item runs the whole grid against a fixed seed: per-cell PNGs with the full recipe embedded, a labeled contact sheet, sweep_summary.csv, and a JSON manifest. It turns sampler folklore into a spreadsheet.

    Grid mode vs combos mode

    There are two ways to tell it what to run, and mode picks between them.

    Grid is the default: fill any axis field and the node takes the cross product. The ones you'll actually set first:

    • samplers - all 13 ship pre-populated (euler, res_2m, res_2s, deis_3m, abnorsett_3m/4m, lawson4_4s, etdrk4_4s, er_sde, rk6_7s, lcm, the hybrids). Delete what you don't want; blank = not swept.
    • schedulers - built-in curves and saved ★ sigma-preset names mixed freely; each entry resolves as a schedule first, then as a sigmas preset. Unknown names are skipped with a console warning, never an abort.
    • steps, windows (e.g. 5-12, 6-12 → start/end step), etas, noise_types, cfgs, shift_modes/mus, hybrid_steps, megapixels.

    pair_lock zips samplers+schedulers into fixed pairs for testing a known recipe; off = full cross product. max_combos (default 64) is a hard cap - over-cap plans refuse with an axis-size breakdown rather than silently truncating, which is the honest behavior you want.

    Combos mode skips the grid: one JSON object per line in combos_json (or a file via combos_file), using the same widget keys as Ultra presets and PNG recipes. {"s2_sampler": "lcm_hybrid", "s2_steps": 12, "s2_start_step": 6} - copy-pasteable straight off a stamped image. Campaigns become versionable files.

    The knobs that shape a sane sweep

    stage_scope (s1 / s2 / s3 / all_stages) decides which stage the swept keys hit, and the workflow intent is spelled out in the tooltip: perfect Stage 1 first, then sweep s2/s3 refinement on the short list. quick_screen forces S1-only at screen_megapixels for cheap wide grids - and for s2/s3-only sweeps, the post-S1 latent is computed once and reused across cells, which is what makes a 40-cell refine sweep finish in minutes.

    The newer axes are where this gets genuinely interesting for Krea 2's ecosystem: lora_mode=bakeoff tests each style LoRA against the bare checkpoint (the lora_triggers field even merges each LoRA's trigger words into its cell's prompt), prompts adds a prompt axis (slowest, deliberately), and sigma_profiles sweeps saved sigma shapes as full per-cell profiles. Failed cells never kill a sweep - red tile plus a manifest error - and interrupts finalize the files on disk before the queue stops.

    Wiring and install

    enabled (default on) means "off = the node outputs nothing and Ultra runs a normal single generation" - you can leave the sweep wired in and toggle it. sweep_preset backs onto sweep_presets.json; selecting one writes its values into the panel so you always see exactly what will run, and editing any field flips it back to custom.

    It ships with the rest of the pack:

    cd ComfyUI/custom_nodes
    git clone https://github.com/EricRollei/Krea2_ComfyUI_Advanced
    

    Restart, and remember the pack needs a diffusers build with Krea2Pipeline plus Krea 2 weights - see the pack README for the --force-reinstall install from the HF git repo.

    Where people get burned

    Watch out for axes that differ only by inert knobs - the plan dedupes cells where eta/noise don't apply (e.g. on euler/lcm), but it can't read your mind about whether sweeping megapixels is worth the VRAM. Resolution changes the noise grid, so composition shifts with size even at a fixed seed; that's part of what you're measuring, but it makes megapixels sweeps noisier to read. Start wide with quick_screen, narrow with stage_scope, and let Sweep → Preset promote the winners.

    CategoryEric/Krea2

    Inputs (45)

    NameTypeDefaultDescription
    modeCOMBOgrid2 options: grid, combos
    stage_scopeCOMBOs1Which stage(s) the swept keys apply to. all_stages sets the same value on s1+s2+s3 together (NOT a per-stage cross product). Workflow intent: perfect S1 first (quick_screen), then sweep s2/s3 refinement on the short list.
    samplersoptSTRINGComma/newline-separated. Allowed: euler, res_2m, res_2s, deis_3m, abnorsett_3m, abnorsett_4m, lawson4_4s, etdrk4_4s, er_sde, rk6_7s, lcm, lcm_hybrid, lcm_hybrid2. Starts empty (2026-10-16); the preset 'All samplers + schedulers (lists only)' fills in every option. Blank = not swept. Crossed with hybrid_steps, only the hybrid samplers expand per split value; other samplers keep one cell (inert-split dedupe).
    schedulersoptSTRINGBuilt-in schedules AND saved sigma-shape ★ preset names, mixed freely: each entry is tried as a schedule (linear, balanced, karras, beta57, beta, bong_tangent, exponential, linear_quadratic, simple) first, then as a sigmas preset (applied as a full profile for that cell); unknown names are SKIPPED with a console warning instead of aborting. Starts empty; the preset 'All samplers + schedulers (lists only)' fills in everything. Blank = not swept.
    pair_lockoptBOOLEANfalseON: zip samplers+schedulers line-by-line into fixed pairs (lists must be equal length) - for testing known recipes. OFF: full cross product.
    stepsoptSTRINGe.g. '8, 10, 12'. Maps to <stage>_steps.
    windowsoptSTRINGRefine windows 'start-end', e.g. '5-12, 6-12, 7-12'. Maps to <stage>_start_step/<stage>_end_step. Pair with matching steps or leave steps blank and window within the panel's step count.
    etasoptSTRINGe.g. '0, 0.3, 1.0' -> <stage>_eta.
    noise_typesoptSTRINGAllowed: white, low_freq, high_freq, pink. lcm-family samplers force white (per the 2026-07-23 rainbow fix); those cells are auto-coerced and labeled, never silently divergent.
    cfgsoptSTRINGe.g. '1.0, 1.5, 2.0' -> <stage>_cfg. Diagnostic for RAW-drifted fine tunes: tolerance of cfg>1 tracks how far training eroded the turbo distillation.
    shift_modesoptSTRINGdistilled_shift values to sweep: fixed / resolution / manual.
    musoptSTRINGshift mu values for the scoped stage(s), e.g. '1.15, 2.1, 3.0'. Forces distilled_shift=manual on those cells (unless shift_modes is also swept, which then wins the cross).
    combos_fileoptSTRINGcombos mode: path to a JSON file with the runs - either a JSON array of override objects, {"combos": [...]}, or JSON-lines. Takes precedence over combos_json below. Keeps sweep campaigns as versionable files.
    combos_jsonoptSTRINGcombos mode: one JSON object per line using Ultra V2 widget keys (same keys as presets / PNG recipes), e.g. {"s2_sampler": "lcm_hybrid", "s2_steps": 12, "s2_start_step": 6} Copy-paste directly from a preset or a stamped PNG.
    seed_policyoptCOMBOfixedfixed: every cell uses the Ultra panel seed (required for fair comparison). seeds_axis: 'seeds' below becomes one more grid axis.
    seedsoptSTRINGseeds_axis values, e.g. '1, 2, 3'.
    quick_screenoptBOOLEANfalseForce S1-only at screen_megapixels for every cell (upscale_to_stage2/3 = 0). Cheap wide screening; re-run the short list full-pipeline after.
    screen_megapixelsoptFLOAT1.000.25–8—
    score_metricsoptBOOLEANtrueCompute technical metrics per cell (CPU, ms): Laplacian + Tenengrad sharpness, wavelet noise sigma, flat-region chroma variance (splotch detector), clip fraction. Scores annotate + can sort the sheet; they never hide a cell.
    sort_byoptCOMBOplan_orderContact-sheet order. Metrics sort best-first; time_s fastest-first. Manifest keeps plan order + this ordering separately.
    tile_pxoptINT448128–1024—
    max_combosoptINT00–1000000 = no cap (default since 2026-10-15; the readout under the node shows the cell count as you edit). A number > 0 is a hard cap: over-cap plans REFUSE with an axis-size breakdown (no silent truncation).
    output_folderoptSTRINGWhere sweep folders are created. Blank = ComfyUI output/sweeps. An absolute path is used as-is; a relative path is created under the ComfyUI output directory.
    hybrid_stepsoptSTRINGSweep the hybrids' handoff in EXACT lcm steps, e.g. '3, 4, 5, 6' -> <stage>_hybrid_steps for the scoped stage(s). Only lcm_hybrid / lcm_hybrid2 read it - pair with a hybrid in the samplers axis (or on the panel); inert cells are deduped. Values clamp inside each stage's window so both halves always run; sweeping values past window-1 collapses to the same cell.
    megapixelsoptSTRINGSweep s1_megapixels, e.g. '0.5, 1.0, 2.0, 4.5'. Always targets Stage 1 regardless of stage_scope (stage 2/3 sizes chain from it via the upscale factors). Mixed cell sizes are fine: metrics and the contact sheet use each cell's native output; the IMAGE batch output center-pads smaller cells to the largest. quick_screen does NOT override a swept megapixels value. Note: resolution changes the noise grid, so composition shifts with size even at a fixed seed - that is part of what you are measuring.
    lora_modeoptCOMBOnonebakeoff: each LoRA in lora_files at each lora_strengths value - by default the swept LoRA runs ALONE (panel stack replaced) so every style is measured against the bare checkpoint; keep_panel_stack layers it on top instead. strength_ladder: keep the panel-declared stack, step ONLY the lora_target entry through lora_strengths. stacks: one whole stack per lora_files line ('comic @0.8 | V0id @0.6') to test LoRAs together. Sweep-injected entries are ephemeral (fused per cell, cleaned per cell).
    lora_filesoptSTRINGbakeoff: one LoRA per line - a name from the loras folder (same resolution as the LoRA Stack node), a full path, or just the file name (found in any subfolder). stacks: one STACK per line, LoRAs separated by | , each 'name @weight' or 'name @s1/s2/s3' (no @ = the LoRA Catalog card weight, else 1.0), e.g. comic_book_v2 @0.8 | V0id @0.6 PICKS - 'random:' draws xN, 'all:' takes every one - from a folder or the LoRA Catalog: random: graphic-art x6 (a loras/Krea2 subfolder, a loras-folder path or a full path) random: graphic-art x4 @0.8 | V0id @0.6 (stacks: 4 stacks, a new pick each) Parkes | all: illustration (stacks: Parkes with each other LoRA, once) cross: random: illustration x4 | random: graphic-arts x4 (stacks: every combination) random: catalog comic x4 | partner: best (stacks: each pick + its best-rated partner from your stacks, else what Civitai users stack it with; partner: random = any of your good partners; @w and seed= work) random: catalog watercolor x4 (catalog: every word on the card - name, file, folder, kind, effect, facets, artists, triggers, creator) CATALOG FILTERS (combine freely): tag="art nouveau" creator=Adel_Ai (or author=; case, _ and - ignored) kind=style_medium role=style arch=Krea2 (the default; Flux, Z_Image, Wan ... or any) e.g. random: catalog creator=Adel_Ai tag=watercolor x4 PER PROMPT (needs prompts; must be the only line): every prompt gets its OWN pick, paired instead of crossed with all prompts. bakeoff: random: catalog creator=Adel_Ai per prompt (one LoRA per prompt; 'x2 per prompt' = two single-LoRA cells per prompt) stacks: random: catalog creator=Adel_Ai x2 per prompt @0.8 (ONE stack of 2 per prompt; add '| V0id @0.6' for a fixed LoRA in every stack; one 'per prompt' anywhere in the line covers every random slot, e.g. 'random: comic | random: illustration per prompt'; include_baseline / stack_parts work within each prompt) SEEDS: none = the same picks every run; seed=7 = another repeatable draw; seed=new = a fresh draw every queue. Show plan lists the picks; every cell records them. WORD OPTIONS (card > Prompt words > Word options): 'name #option @0.8' = that option's words; 'name #each' = one cell per option; 'name #random' = one option drawn at random (repeatable; on a 'per prompt' line every pick draws its own). Lines starting with # are ignored.
    lora_strengthsoptSTRINGcarde.g. '0.6, 0.8, 1.0'. One value drives all three stage weights (weight_s1=s2=s3). bakeoff: every file at every value; ladder: the target entry's values. stacks: optional multipliers - every stack (and every single-LoRA baseline) runs once per value, all its weights scaled (blank = the weights as written). bakeoff also takes card / card_low / card_high: each LoRA at its LoRA Catalog card weight (your editor weight wins; no card = 1.0), mixable with numbers, e.g. 'card, 1.0'.
    lora_targetoptSTRING1strength_ladder: which panel-stack entry to vary - 1-based index, or a name substring (matches lora_name/filename).
    include_baselineoptBOOLEANtrueAdd one reference cell: bakeoff = no LoRA at all (or the panel stack alone when keep_panel_stack is ON); ladder = the panel stack exactly as declared. stacks = that reference cell PLUS every distinct LoRA of the lines alone at its line weight, so each combination can be judged against its parts.
    keep_panel_stackoptBOOLEANfalsebakeoff and stacks: append the swept LoRA (or stack) ON TOP of the panel stack instead of replacing it.
    promptsoptSTRINGOne prompt per line = a prompt axis (slowest axis, so the sheet and CSV read in prompt-major blocks). Each prompt cell re-encodes (prompt_conditioning is cleared; ~seconds/cell). Metric comparisons are only meaningful WITHIN a prompt block. PROMPT FILES (one prompt per line; blank and # lines skipped): file: D:\prompts\portraits.txt every prompt in the file (all: works too) random: D:\prompts\portraits.txt x20 20 different prompts drawn at random seed=7 = another repeatable draw; seed=new = a fresh draw every queue A relative path starts from the ComfyUI folder. Mix freely with typed prompts. Every cell records the actual prompt text; editing the file rebuilds the plan.
    sigma_profilesoptSTRINGComma/newline list of saved sigma-shape preset NAMES (the ★ presets of the Eric Krea2 Sigmas node). Each cell rebuilds the full bundle from that preset and it OVERRIDES any wired sigmas node for that cell. Blank = not swept.
    lora_triggersoptCOMBOsaved pickWords added to each cell's prompt for its swept LoRAs (bakeoff and stacks; baseline cells get none). Check them with 'Show plan' before you queue. none = nothing. trigger = each LoRA's main trigger from its card (can be a few words). Enough on the base model, often too weak on fine-tunes. full = each LoRA's whole recipe: official words + trigger + style phrase. saved pick = the words picked for the loaded checkpoint (yours, else a trigger test's), else trigger + style phrase. Best default. full and saved pick merge a stack's words into one tidy list. Details (older names): bakeoff and stacks: merge each cell's swept-LoRA trigger words into that cell's prompt (from the LoRA Stack node's cached trigger database; first-seen files may do one network lookup, non-fatal on failure). Only sweep-injected entries contribute - baseline and ladder cells are untouched. A cell whose prompt text changes re-encodes automatically. Effective prompt + triggers are recorded in the cell's metadata. catalog: ONE trigger per swept LoRA - its LoRA Catalog card's first active trigger (your priority ones first, switched-off ones never), else the first cached Civitai trigger - always PREPENDED (Krea 2 follows a leading trigger best). recipe (2026-10-15): each swept LoRA's LoRA Catalog prompt words for this checkpoint - your pick, else a trigger test's pick, else trigger + style phrase - prepended; an empty pick adds nothing.
    enabledoptBOOLEANtrueOff = the node outputs nothing and Ultra V2 runs a normal single generation - leave the sweep wired in the workflow and toggle it.
    sweep_presetoptCOMBOcustomNamed sweep plans from sweep_presets.json. Selecting one WRITES its values into this panel (so you can see exactly what will run); editing any field flips back to 'custom'. Save the panel as a preset with the ★ button. Headless/API: the named preset applies fully. Ask Claude to author experiment presets.
    preset_notesoptSTRINGOne or two sentences: what this sweep preset is FOR and what it does. Saved and loaded with the ★ preset - so future-you remembers why it exists.
    run_notesoptSTRINGWhy you are running THIS sweep today (what you are checking, what changed since last time). Written into the sweep folder - sweep_manifest.json and sweep_info.txt - with the preset name and preset notes. Never saved into a preset, and typing here does not flip the preset dropdown to custom.
    share_sheetsoptCOMBOstripsAfter the sweep finishes, write ready-to-post JPEGs into <sweep>/share/. strips = one labelled strip per prompt (768 px tiles, wraps into more rows, max 6000 px a side), with the settings on top and each LoRA's short name, weight, creator and Civitai model number under its image. singles = each image with its own labels. To add star ratings, make them again from the catalog's Sweeps page after you rate.
    share_backgroundoptCOMBOblackBackground colour of the share sheets.
    stack_partsoptCOMBOeach LoRA alonestacks mode with include_baseline ON - the cells added next to the stacks: each LoRA alone: the no-LoRA reference + every LoRA of the lines alone (as before). reference only: just the no-LoRA reference - screen many stacks fast, e.g. 'cross: random: illustration x4 | random: graphic-arts x4 | random: treatment x4' = 64 stacks + 1 reference. build up: the reference, then each stack one LoRA at a time in the order written: A, A+B, A+B+C - what each layer adds. leave one out: the reference + each stack minus one LoRA at a time (A+B, A+C, B+C) - which LoRA a stack can do without. Repeated cells are made once. include_baseline OFF = no extra cells at all.
    sde_bandsoptSTRINGSweep the SDE band as 'hi-lo' sigma pairs, e.g. '1-0, 0.93-0.44, 1-0.5, 0.5-0.15' -> <stage>_sde_hi / <stage>_sde_lo for the scoped stage(s). Churn only happens on steps whose next sigma is inside the band, so pair with eta > 0 (etas axis or panel). Cells where the band is inert (euler / lcm / er_sde, or an explicit eta of 0) are deduped. Blank = not swept.
    churn_formsoptSTRINGSweep the churn form: 've, rf' -> churn_form (all stages). ve = previous noise injection, rf = rectified-flow form. eta values mean different things in the two forms, so cross with an etas axis. Blank = not swept.
    hybrid_handoffsoptSTRINGSweep the lcm-hybrid handoff in sigma, e.g. '0.8, 0.7, 0.6, 0.5' -> <stage>_hybrid_handoff for the scoped stage(s). The lcm head runs until the ladder reaches that sigma (first step at or below it). Overrides hybrid_steps for those cells; inert on non-hybrid samplers (deduped). Blank = not swept.
    trigger_joinsoptSTRINGTEST AXIS (temporary): 'comma, period' - how the swept LoRAs' trigger words join the prompt (comma = 'trigA, trigB, prompt'; period = 'trigA. trigB. Prompt'). Needs lora_mode bakeoff or stacks with lora_triggers on; baseline cells are deduped. Only the swept LoRAs' words - an upstream stack's words were already joined by its own trigger_join. Blank = not swept.

    Outputs (1)

    NameTypeDescription
    sweepKREA2_SWEEP—