ALMA.
Guides

Story & Story v2 — best practices

Tuned from real QA reviews. Pick the combo that fits your beat, lock your character design, and keep lettering bold. Generate in Discord with /story or /story-v2.

Best practices — /story

Comic strip with captions, delivered as one image (panels are no longer attached separately). Attach a clear character ref and prefer 4 panels.

Always

  • Character ref: one clear full-body / ¾ shot, plain background, final outfit — the art model sees this photo every panel.
  • Use Enhance on vague premises (“make her look cool”) so DNA + beat progression are locked before generate.
  • Lock only what is in the ref: hair, eyes, sleeve length, coat/jacket, tie, present accessories. Prefer “no gloves” over “bare arms” if the shirt has sleeves — sleeveless rewrites are a common fail.
  • Name what drifts — models invent cat ears/gloves, strip coats, turn eyes into glowing voids, and redesign hair mid-strip.
  • Captions: short voice (≤8 words) or omit — not “surveys the battlefield” art-restating. Strip captions sit in a fixed bottom bar; never paint text in panels.
  • Props + foes need setup: summon/draw the sword before the clash; establish the army before it charges.
  • Prefer 4 panels (draw → cast → clash → win); at 6, outfit drift and pose spam rise.
  • Never paint text in panels — captions are added after.

Gemini 3.1 Flash Lite

6 panels writer: Grok 4.20

Dramatic anime key-art: sharp lines, glowing magic accents, cinematic lighting, consistent palette across panels

  • Recent go-to art tier — scored ~8 with Grok 4.20 captions and a single character ref.
  • Its main slip is outfit drift late in the strip (sleeve length, skirt hem, thigh-highs) and eyes going purple — restate the exact outfit + eye color.
  • For 6 panels, add a clear transition beat (e.g. summoning) so the monster/weapon reveal is not abrupt.

Example premise

Six-panel duel; keep the reference heroine identical each panel (silver hair, cat ears, blue eyes, hair clip, scalloped strapless black dress, choker). Show her draw the magic sword before the clash; resolved victory.

Seedream 4.5 (uncensored)

4 panels writer: Grok 4.20

High-energy anime fantasy, crisp ink, cel shading, readable faces/weapons, shared lighting across panels

  • Strong with character refs — still name coat length, trim, outfit colors in the premise.
  • Writer: Grok 4.20 (not lightweight code models) for tighter DNA repetition.
  • 4 panels preferred; 6 scored OK but outfit drift rose.

Example premise

Two mage-swordswomen defend a ruined arena; white-haired lead in long black coat with fur collar, pink-haired partner in black outfit — keep designs identical every panel; they win with a dual slash.

Grok Imagine Quality (uncensored)

4 panels writer: Grok 4.20

Cinematic anime/manga lighting, consistent palette, sharp silhouettes, heroic poses

  • 6-panel magic-sword runs still score ~4–5.6: coat/sleeve flips and art-restating captions. Prefer 4 panels + Enhance; lock shirt sleeves + jacket every panel from the ref.
  • Put weapons in DNA or a draw beat; setup foes before the charge; captions = short voice or empty — not “one strike ends the army”.
  • Never paint captions in the art — the strip bar is the only lettering.

Example premise

Four-panel rise. Keep ref exact (hair, yellow eyes, white shirt + black tie + black jacket on shoulders, no gloves). Draw conjured sword → foe appears → clash → win. Short voice captions or none.

Gemini 2.5 Flash

4 panels writer: Grok 4.20

Clean comic illustration, readable action, consistent character sheets

  • Good default cost/quality; keep style text short and shared; attach a character ref (panels now see the photo).
  • 6-panel gothic reviews (~5.6): teddy bow, bat-wing clips, choker, pink underskirt, and eye color drifted — name each accessory + exact colors in the premise.
  • Prefer 4 panels + one shared background mood; do not paint dialogue/captions in the art.

Example premise

Four-panel rise: stance → cast → clash → victory. Keep ref exact (amber eyes, twin-tails + lace bows, bat-wing clips, black choker, dress + pink underskirt, brown teddy with blue bow); one gothic mood throughout.

Other art models fallback

4 panels writer: Grok 4.20

One shared style line; crisp subjects; no in-image text

  • Start with 4 panels + Grok 4.20 writer + character attachment.
  • If consistency slips, shorten the cast and lock colors/outfit in the premise.

Example premise

Hero faces a single clear threat across 4 panels; identical costume every panel; punchy captions that match the art.

Best practices — /story-v2

Lettered page (bubbles baked on). Use Simple to pick a curated combo (layout + style), or Complex to tune packs/writers. Attach a character ref.

Always

  • Simple combos: Action page · Quad grid (4 boxes) · Cinematic rows · Impact splash · Webtoon · Western color — all 4–5 panels on Grok 4.20 stages.
  • Fights / magic-sword premises: prefer Action page or Impact splash. Cinematic rows are for mood — they drift into hair/skirt close-crops and hide the action.
  • Use Enhance on vague premises so Character DNA + beats lock before generate.
  • Prefer 4–5 panels; 6 only for dense fight lists.
  • Lock only what is in the ref every panel — hair style (twin-tails vs straight), tee graphic, plaid/skirt, choker; do not invent ears/gloves; keep full color (manga + colored refs no longer force grayscale).
  • Lettering is the weak spot: keep speech ≤~6 short words / shouts shorter; enlarge and bold SFX; high-contrast captions in gutters; pad bubbles off faces/hair.
  • Fill or remove empty balloons — no blank boxes, no mid-word cutoffs (“HYPERDIMENSI…”, “unbrea…”). Prefer speech over tell-y narration (“I am the strongest”, “crowd is impressed”).
  • Props need origin: sword/loot sack/feast → draw, pick up, or order/serve beat first. One foe with setup→clash→payoff; merge duplicate eat/win panels.
  • Every panel needs a readable face or clear action — avoid empty finale panels and fabric-only crops.
  • Character ref: clean full-body; style_image optional and matching the pack.

Gemini 3.1 Flash Lite

mode: simple 4 panels manga action_stack bubbles: normal writer: grok-4-20 (script · layout · safe-zone)
  • Current default art in reviews — scored 7.5–8 on Simple with Grok 4.20 stages.
  • Consistency is strong; push it further by naming cat ears + blue X-clip in every panel and locking the scalloped neckline.
  • Boost SFX size/contrast and bubble outline weight; do not leave empty balloons or blank boxes.

Example premise

Make the reference heroine the strongest magic-and-sword fighter; keep her design exactly (silver hair, cat ears, blue eyes, X-clip, black scalloped dress, choker). Cool arena victory page; bold readable SFX.

Seedream 4.5 (uncensored)

mode: simple 4 panels manga action_stack bubbles: normal writer: grok-4-20 (script · layout · safe-zone)
  • Scored ~8 with Simple + manga + 4 panels + Grok 4.20 stages.
  • Watch bubble contrast on glowing effects; keep speech off ornate clothing.
  • Finish every balloon (no empty tails).

Example premise

Make these characters the strongest magic-and-sword fighters; keep designs exactly as the reference; cool arena victory page.

Grok Imagine Quality (uncensored)

mode: simple 4 panels manga action_stack bubbles: normal writer: grok-4-20 (script · layout · safe-zone)
  • Strong with ≤3 `input_references` — still follows DNA text over the photo if they conflict; never paste a different colorway in the premise.
  • Lock sleeves/coat/hair from the ref every panel; setup one foe before the clash; one victory panel only; keep balloons very short and complete.
  • Prefer 4–5 distinct beats; pad bubbles off faces; Western/Webtoon packs if you want punchier color comics.

Example premise

Four-panel fight. Keep ref exact (hair, eyes, white shirt + black tie + jacket). Short speech or silent action panels. Draw weapon → foe setup → clash → one win; full color, no empty balloons.

Gemini 2.5 Flash

mode: simple 4 panels western dialogue_2x2 bubbles: normal writer: grok-4-20 (script · layout · safe-zone)
  • Cheaper art tier — the Western color combo (4-box 2×2) suits it well.
  • Keep panel count at 4 and writers on Grok 4.20; short dialogue.

Example premise

Four-box lettered page: declare → clash → turn → victory. Match character ref; short dialogue; bold outlined bubbles.

Other art models fallback

mode: simple 4 panels manga auto bubbles: normal writer: grok-4-20 (all stages)
  • Use Simple and pick the combo that matches the beat (grid for dialogue, splash for a money shot).
  • Safe-zone model must support vision (· img).

Example premise

One heroine, one conflict, clear ending; lock hair/eyes/outfit; 4 panels; bold readable lettering.

Best practices — /story-v3 (Alma Page)

Identity-sheet comic page: beat template → storyboard confirm → sheet → silent panels → ALMA lettering (safezone places bubbles). Character ref required. Continue reuses the sheet.

Always

  • Always attach a clear character ref — Alma Page builds an identity sheet from it and reuses it on every panel + Continue.
  • Do not paste a Character DNA block that contradicts the photo (e.g. lavender OC DNA + Makima ref). With a ref, premise DNA is stripped and the photo owns identity.
  • One character ref = one designed face. Foes stay faceless silhouettes unless you attach a second character ref. Vague premises (“War hero”) → use Enhance.
  • Pick a beat template: Kishōtenketsu (mood/slice) or Action arc (fights). Fight premises force Action arc even if Reveal was selected.
  • After Generate you get a storyboard — Confirm before image spend. Edit dialogue there; cancel refunds.
  • No layout AI: script text is canonical; safezone vision only picks bubble placement (faces avoided).
  • Keep speech short (≤~6 words). Dialogue/reaction beats must have real bubble text — never `bubbles:[]` with a dialogue label (empty balloons kill lettering scores).
  • Every page needs ≥1 real spoken line (not SHING/CLANG-only) plus optional SFX. DNA must list concrete hair/eyes/outfit from the photo — not a generic placeholder.
  • Weapons used in later panels must be in DNA or an earlier draw/grip beat — Action arc panel 2 is the draw slot.
  • Lock full outfit in DNA (including bottoms/coat state). “More skin” without DNA bottoms = drift. Sheet + ref win over scene adjectives.
  • Prefer ≤1 close-up per page; avoid hair/cleavage/hand-only crops.
  • Fights: Action arc names the foe in panel 1, draw/summon before clash, one payoff. Romance/mood: Kishōtenketsu (not Action stack).
  • Prefer models that honor multi-image `input_references` (Gemini / Seedream tiers). Sheet + char ref both go on panel gens.
  • House QA: heuristics logged; vision + writing judges run on every job while Alma Page is in A/B.

Gemini 2.5 Flash

mode: alma 4 panels manga dialogue_2x2 bubbles: normal writer: grok-4-20 (script) · vision safe-zone
  • Good default for Alma Page cost/latency with multi-ref.
  • Use Kishōtenketsu for mood pages; Action arc for fights.
  • Confirm storyboard before spend — fix weak dialogue there.

Example premise

Four-beat fight. Keep ref exact. Action arc: name the foe → draw sword → clash → one win. At least one short line or SFX; Action stack layout.

Gemini 3.1 Flash Lite

mode: alma 4 panels manga action_stack bubbles: normal writer: grok-4-20 (script) · vision safe-zone
  • Strong ref adherence for sheet + panels.
  • Lock sleeves/coat from the photo; prefer “no gloves” wording.

Example premise

Kishōtenketsu slice: introduce heroine → develop tension → twist → resolve. Match character ref every panel.

Seedream 4.5 (uncensored)

mode: alma 4 panels manga action_stack bubbles: normal writer: grok-4-20 (script) · vision safe-zone
  • Good with character sheets; keep dialogue short for lettering.

Example premise

Action arc arena duel; identity from ref; silent clash panel OK.

Grok Imagine Quality (uncensored)

mode: alma 4 panels webtoon action_stack bubbles: normal writer: grok-4-20 (script) · vision safe-zone
  • Honors refs but DNA text still wins on conflict — never paste a different hair/eye/outfit colorway in the premise.
  • Max 3 input refs: sheet + character is enough; Action arc for trench/fight pages.
  • Fill speech/SFX in script; model must not paint empty balloons (Alma letters after).

Example premise

Action arc trench charge. Match character ref exactly (no lavender recolor). Draw weapon → cover → clash → payoff. Short lines or SFX.

GPT Image 1 Mini

mode: alma 4 panels manga action_stack bubbles: normal writer: grok-4-20 (script) · vision safe-zone
  • Draft/cheap tier — drifts cast and paints empty boxes more than Gemini. Alma now sends quality=high + input_fidelity=high with refs.
  • One ref only: keep foes as silhouettes; attach a second character ref if you need a designed rival.
  • Prefer Gemini/Seedream for scored A/B pages; use Mini for cost previews.

Example premise

Action arc. One heroine from ref; faceless enemy silhouettes; face readable every panel; no blank caption boxes.

Other art models fallback

mode: alma 4 panels manga auto bubbles: normal writer: grok-4-20 script + vision safe-zone
  • Prefer models with maxReferenceImages ≥ 2 so sheet + character both attach.
  • Use storyboard Confirm to avoid bad-script image spend.

Example premise

One heroine, clear foe setup, twist, single payoff; short speech.

Simple combos (Story v2)

Simple mode keeps panel counts low (4–5) but offers a spread of layouts and styles. Pick one in the review panel.

Action page

4–5 panels · manga · dynamic action stack · Grok 4.20 writers

Best all-round fight page: manga ink + action stack reads clean and scored ~8 with character refs.

Quad grid (4 boxes)

4 equal boxes on one page · manga · dialogue 2×2

Classic even 2×2 grid — four boxes on one page instead of a vertical strip. Great for dialogue / reaction beats.

Cinematic rows

4–5 wide letterbox rows · manga · normal bubbles

Widescreen stacked rows for mood / establishing beats. Prefer Action page for fights — cinematic rows need readable medium shots + high-contrast captions.

Impact splash

Big hero splash + support panels · manga

One dominant splash for the money shot; keep the cast small so the hero is not buried in particle soup.

Webtoon column

Tall vertical scroll · full-color webtoon · Grok 4.20 writers

Vertical color storytelling with soft cel shading — reads well on phones.

Western color

4 boxes · saturated western color · dialogue 2×2

Bold superhero-comic color with a clean 2×2 grid — a good pick for cheaper art models.

Make a page.

Open Discord and run /story-v2 — the review panel has these combos built in.

Back home