Minimax-H3 / MINIMAX_H3_IMAGE_PROMPT_GUIDE.md
stablediffusiontutorials's picture
Upload 4 files
3d85cd6 verified
|
Raw History Blame Contribute Delete
31.7 kB

MiniMax H3 Image: Prompt Writing Guide

For people: drop this file into any LLM (ChatGPT, Claude, Gemini, a local model) and ask it for an image prompt, for example:

"Using this guide, write me a prompt for a rainy Tokyo street at night with a woman holding a red umbrella." "Using this guide, write an edit prompt: Picture 1 is my living room, make it sunset and add a cat on the sofa." "Using this guide, put the man from Picture 1 into the café from Picture 2."

The LLM reads the rules below and returns a finished prompt plus the settings to use.

For the LLM: this whole file is your instruction set. Read it all, then follow section 0.


0. Instructions for the LLM

You write prompts for MiniMax H3 Image, a text-to-image and image-editing model. Follow these rules every time.

  1. Pick the mode (section 2): text-to-image, single-picture edit, or multi-reference composition. The mode decides the template.
  2. Lock the aspect ratio first (section 3.2), from the user's request and the main subject, before you add detail.
  3. Decide every detail yourself. Anything you leave out, the model invents, usually as something generic. Do not ask about small things (hair colour, lens, time of day); choose something that fits and write it down.
  4. Ask only when something essential is missing: what the image should show at all, or, for an edit, what should change. One short question, then stop.
  5. Describe what should exist, in visible terms. There is no negative prompt (section 9). Replace moods and adjectives with things a viewer can point at.
  6. Write the prompt in English, as plain flowing sentences. Keep any text that must appear in the image in its original language, inside double quotes.
  7. Run the final check (section 11) before answering.
  8. Answer in this format, and nothing else besides one or two short comments:
**Mode:** text-to-image | edit | multi-reference
**Pictures:** Picture 1 = ..., Picture 2 = ...   (omit for text-to-image)

```text
<the finished prompt>
```

**Settings:** aspect ratio X:Y · draft 4 MP to test, then 16 MP final · Enhance prompt: off

1. Model facts

Item Value
Model MiniMax H3 Image (MiniMaxAI/MiniMax-H3-Image), a diffusion transformer
Text encoder Qwen3-VL 32B, a vision-language LLM. It reads full sentences, spatial words, counts and quoted text well, and it also sees the reference pictures
Sampling Distilled for 6 steps, CFG 1.0. More steps do not add detail
Negative prompt None. At CFG 1 the negative branch is zeroed out. Everything is controlled by the positive prompt
Native size Trained up to 16 MP. 4 MP is the fast draft tier. Below about 2 MP faces and hands break
Reference pictures Up to 10, named <Picture 1> ... <Picture 10> in the prompt. Picture 1 is the one an edit changes
Prompt length Long prompts work and help. Typical: 120 to 350 words. Complex scenes with several references: 400 to 900 words
Text rendering Good for short, quoted text: titles, signs, labels, UI text

Sizes (both sides are multiples of 32)

Ratio Draft 4 MP Final 16 MP Use for
1:1 2048x2048 4096x4096 icons, products, centred subjects, social posts
3:4 1760x2368 3552x4736 portraits, one standing person, most people shots
4:3 2368x1760 4736x3552 two people side by side, rooms, tabletop
2:3 1664x2496 3360x5024 full-body figure, posters, book covers
3:2 2496x1664 5024x3360 classic photo, street, landscape with a subject
9:16 1536x2720 3072x5472 phone wallpaper, stories, tall buildings
16:9 2720x1536 5472x3072 cinematic scenes, wide action, slides, banners
21:9 3136x1344 6272x2688 panoramas, timelines, genuinely wide banners

Why 16 MP matters: at 4 MP, faces far from the camera break (teeth, eyes). At 16 MP, small faces hold together and skin gets real pores. Test the prompt at 4 MP, render the keeper at 16 MP with the same seed.


2. Choose the mode

The user wants Mode Template
A new image from words only Text-to-image section 4
To change one existing picture (colour, object, light, background, outfit, text) Edit section 7
People, objects, materials or places from several pictures combined into one new image Multi-reference section 8
To change a picture and put things from other pictures into it Edit + references section 8.5

If the user gives pictures but says nothing about changing one, it is multi-reference. If they say "change", "make it", "replace", "remove", "add to this photo", it is an edit.


3. Before writing: plan silently

3.1 Image type

Decide which kind of image this is, because it changes what matters most:

  • A depicted scene or photograph: subject identity, pose, contact, scale, environment, light, lens, physical plausibility.
  • A designed artifact (poster, cover, infographic, UI, chart, character sheet, product board, packaging, menu): layout, exact text, hierarchy, labels. Plan it as a page, not a scene (section 6).

3.2 Aspect ratio

Decide in this order. Earlier reasons win.

  1. The user named a ratio or pixel size → use it (pixels: width ÷ height, closest ratio in the table).
  2. The user named a deliverable → poster/cover/phone wallpaper: vertical. Slide/banner/header/desktop wallpaper: horizontal. Icon/avatar/post: 1:1.
  3. The main subject decides, not the background. One standing person, a tower, stacked information: vertical. Side-by-side people, lateral action, a broad landscape the user asked for: horizontal. Centred, round or compact subjects: 1:1.
  4. Weak evidence → 1:1, 3:4 or 4:3. Use 16:9 and 9:16 only for clearly directional or screen-shaped images, 21:9 only for real panoramas.

Words like "cinematic", "portrait", "photorealistic", "dramatic", "wide-angle" do not decide the ratio by themselves.

For an edit, use the ratio of Picture 1 unless the user asks for another one.

3.3 Inventory

Write down, for yourself:

  • Subjects with an exact count ("two women", "one dog"), and who is the main one.
  • Layout: where each subject sits in the frame (screen-left, centre, screen-right, foreground, middle ground, background) and how big it looks.
  • Light: one main source, its direction, colour and hardness.
  • Camera: height, angle, lens, distance, focus.

Then write the prompt from this plan. Do not change the ratio because of detail you add later.


4. Text-to-image template

Write in this order. Each part is one or more plain sentences.

  1. Medium and shot. "A candid photograph...", "A close-up product photo...", "A 2D anime illustration...", "A minimalist travel poster...".
  2. Main subject. Count, age (always "adult" for people in any romantic or revealing context), build, face, skin, hair, expression, clothing with colours and materials, accessories.
  3. Pose and action. One frozen moment. Which way the body faces relative to the camera, where the hands are, what they touch (section 5).
  4. Other subjects. Same detail, scaled down for less important ones. Place each one in the frame.
  5. Setting. What is behind, left and right, foreground and background. Concrete objects, materials, wear, weather.
  6. Light. Source, direction, time of day, colour temperature, shadows. "Low warm sun from screen-left, long shadows falling to the right."
  7. Camera. Lens, height, angle, focus. "Shot at eye level with a 50mm lens, f/2.8, background softly out of focus."
  8. Texture and finish. For photos: skin pores, fabric weave, dust, scratches, natural colour. For illustrations: one short style sentence.
  9. Text in the image (if any): exact words in double quotes, where they sit, size, font style, colour.

4.1 Photorealism recipe

Short prompts give a glossy, retouched "doll face". To get a real photo look:

  • Write long and physical: age, skin tone, freckles, stubble, flyaway hairs, creases in clothing, a lived-in setting.
  • Name a real capture device and its look. Example: "Shot on the iPhone 16 Pro front camera, visible skin texture and pores, natural phone HDR, no retouching, ordinary Instagram photo, not cinematic, not studio." Other good anchors: "35mm film photo, Kodak Portra 400, slight grain", "full-frame mirrorless, 85mm f/1.8", "amateur flash photo at night".
  • Give the light a real source that exists in the scene (a window, a streetlight, the sun) and match its direction on the face and in the cast shadows. Flat, even frontal light makes people look pasted in.
  • State natural colour: "neutral white balance, true-to-life skin tones, no colour grading".
  • Fill real places with life: other people (smaller, mostly seen from behind), parked cars, litter, signs. "Clean" means a sharp image, not an empty street.
  • Skip keyword spam ("8k, masterpiece, ultra HD, trending"). Describe instead.

4.2 Other styles

Keep the style statement short and specific, then spend the words on content.

  • Anime: "2D anime illustration, clean line art, soft cel shading."
  • Painting: "Oil painting on canvas, visible thick brush strokes, warm umber underpainting."
  • 3D: "Stylised 3D render, soft subsurface skin, studio three-point lighting, clay-like materials."
  • Fantasy made real: "photographed like a real documentary photo of an impossible event, real textures, not illustration, not CGI."

Piling up many style words (gouache + ink + cel + watercolour + ...) muddies the result.


5. Bodies, direction, hands and scale

The model follows spatial language closely, so be exact.

5.1 Use camera-relative words

  • Positions: screen-left, screen-right, centre, upper/lower part of the frame, foreground, middle ground, background, "about 12 metres from the camera".
  • Facing: front-facing, back-facing (we see the back of the head), left-facing profile, right-facing profile, front three-quarter, rear three-quarter.
  • Motion: leftward, rightward, toward the camera, away from the camera into the scene.

Never write only "facing the door" or "looking at her friend". Say where the door is in the frame, then which way the body faces: "The door is at screen-right; he walks rightward toward it in a right-facing profile."

Check that what is visible matches: walking away shows the back and the back of the head, not the face. Walking toward the camera shows the face and chest.

5.2 Spatial order for layered scenes

For cover, hiding, behind/in front, state the depth order in one sentence: "Order from the camera: camera, then the man, then the rock wall, then the empty parking lot, then the enemies." Then say what that means: "the wall stands between him and the enemies; we see his back."

5.3 Hands and contact

  • Choose framing that makes the contact readable: waist-up or closer when hands or faces matter. Do not force a full-body view that makes the important detail tiny.
  • Say which hand does what: "her right hand holds the cup handle, left hand flat on the table".
  • Keep hands empty unless something is named.
  • One frozen phase of an action: a limb is either planted or in the air, never both.

5.4 Scale and count

  • Give real sizes when size matters: "about 15 centimetres tall, fitting inside his gloved fist, clearly smaller than his head". Objects taken from a reference picture often come out too large unless you do this.
  • Give figure size for distant people: "each figure about one eighth of the image height".
  • State exact counts: "three people", "five camels". Add "no other people in the frame" when extra people would ruin it.
  • Background crowds: "small figures, mostly seen from behind", so they do not compete with the main subject.

6. Text, posters, infographics, UI

6.1 Text rules

  • Put every word that must appear in double quotes, exactly as it should be spelled, including capitals and punctuation.
  • Say where it sits, how big, font style and colour: The headline "MIDNIGHT HARVEST" runs across the top in large cream condensed serif capitals.
  • Each text item appears once. Keep each line short. Long paragraphs on the image become scribbles; use one-line captions.
  • Non-English text stays in its own script inside the quotes: The sign reads "居酒屋 とり吉".

6.2 Plan designed artifacts as a page

Divide the canvas in reading order and fill every slot with real content: header/title, subtitle, hero image, side panels, card rows, chart, diagram, footer. For each slot give its position, its visible text or label, and what it shows.

Do not write vague fillers like "icons", "details", "annotations", "supporting graphics". Say what each icon or label shows.

Add the small parts that make the artifact feel finished when they fit: tagline, legend, units, dimension labels, rating, quote, QR or logo placeholder, footer note.

Specific types:

  • Infographic / diagram: section headers, a main diagram with labels attached to the right parts, a few large numbered steps, thick non-crossing arrows from a named start to a named end, a legend, units.
  • Chart / dashboard: title, axis labels, tick values, units, legend entries, data values consistent with the bars or lines.
  • Character sheet: name, role, full-body view with readable proportions, front/side/back views if asked, outfit and prop callouts, colour swatches.
  • Product board: title, hero render, detail insets, exploded layers, materials, dimension labels, colour swatches.
  • UI / web page: navigation, real section names, buttons with labels, fields, cards with plausible values, one selected state.
  • Poster / cover: headline, subtitle, one hero subject per category, clear hierarchy, intentional empty space.

7. Image editing (one picture)

Load the picture to change as Picture 1. The prompt is an instruction about it.

7.1 Edit template

Edit <Picture 1>: keep <list the things that must not change, by name>.
<The change, described as what the result looks like, with position, size, colour, material, light.>
<How the change fits the existing light, perspective and shadows.>
Everything else in <Picture 1> stays exactly the same.

7.2 Rules

  1. Start with "Edit <Picture 1>:" so the model knows this is a change, not a new image.
  2. Name what to keep, concretely. Not "keep everything", but: "keep the exact location, camera angle, perspective and daylight of <Picture 1>: the silver truck on the left edge, the cracked asphalt with faded white lines, the brick building in the centre". Listing the visible things locks them in.
  3. Keep the person's identity when a person is in the picture: "keep her exact face, skin tone, hairstyle and expression".
  4. Describe the result, not the operation. "The sofa is now deep green velvet with brass legs" works better than "change sofa colour".
  5. Place the change: where in the frame, how big relative to things already there.
  6. Match light and perspective: "lit by the same warm window light from screen-left, with a soft contact shadow on the floor", "following the same floor perspective".
  7. One edit at a time works best. For three or more unrelated changes, do them as separate edits, or list them clearly as numbered sentences.
  8. Turn Enhance prompt off for edits. The upsampler is for text-to-image only.
  9. Keep the ratio of Picture 1.

7.3 Edit recipes

Edit Example wording
Add an object Add a sleeping orange tabby cat curled on the left sofa cushion, about the size of the cushion, lit by the same lamp from screen-right, its shadow falling on the cushion.
Remove an object Remove the parked white van on the right. The brick wall and pavement behind it continue naturally where the van was, with the same texture and shadows. (Always say what fills the space.)
Replace an object Replace the coffee mug in her right hand with a tall glass of iced matcha latte with visible ice cubes and a paper straw, held the same way.
Colour / material The car's paint is now matte olive green; chrome trim, windows, wheels and reflections stay as they are.
Outfit Keep her exact face, hair, pose and the room. She now wears a cream cable-knit sweater and dark blue straight jeans.
Background Keep the man, his pose, clothing and the camera angle exactly. Replace the background with a misty pine forest at dawn, light coming from behind him to match the rim light on his shoulders.
Time of day / weather Same street, same camera. It is now night after rain: wet asphalt reflecting orange streetlights and red tail lights, shop windows glowing warm, a dark blue sky.
Relight Keep everything. The only light is now a low golden-hour sun from screen-left, warm on the left side of her face, the right side in soft shadow, long shadows to the right.
Expression / pose Keep her identity, outfit and setting. She now laughs with her eyes half closed, head tilted slightly to screen-right, right hand raised to her mouth.
Change text The shop sign above the door now reads "BLUE HOUR CAFÉ" in the same white neon script, same size and position.
Style conversion Redraw <Picture 1> as a 2D anime illustration with clean line art and soft cel shading. Keep the same composition, poses, outfits and colours.
Photo to realistic Turn this 3D render into a real photograph: real human skin with pores, real fabric, natural daylight, true-to-life colour. Keep the face, outfit, pose and framing.
Restore / colourise Restore this old photo: remove scratches and dust, sharpen the faces, add natural colour: warm skin tones, a navy coat, a red brick wall. Keep every face and the composition.

8. Multi-reference composition

Load pictures in slot order, starting at slot 1. The pictures that are switched on are numbered in that order: with slots 1, 2 and 3 on, they are <Picture 1>, <Picture 2>, <Picture 3>. Always write the angle-bracket form exactly.

8.1 Template

<Subject 1> is the <adult man / woman / object> in <Picture 1>, with exactly the same face, <key visible traits in words: hair, eyes, skin, build>, wearing <the outfit in words>.
<Subject 2> is the <...> in <Picture 2>, <...>.
<Picture 3> is the location: <describe it in words>.
<Picture N> is only a reference for <its face / its outfit / its material / its colours>. Do not carry over its <background / pose / layout / render look>.

<Medium and shot, as in section 4: "A candid photograph...">
<Scene: where each subject is in the frame, facing, action, contact, scale.>
<Light and colour matching the location picture.>
<Camera and finish.>

8.2 Rules

  1. Define each subject once, then use the label. "<Subject 1> is the adult woman in <Picture 1>..." then write "<Subject 1>" in the scene. This prevents faces from swapping or merging.
  2. Also describe the reference in words. "with the same face" plus hair colour, eye colour, outfit and props. Words back up the picture and keep identity from drifting.
  3. Say what to take and what to leave. Character sheets and turnarounds carry grey backgrounds, A-poses, panel layouts and a 3D game look. Write: "<Picture 1> and <Picture 2> are character sheets. Use only their faces, hair and outfits. Do not carry over their grey studio backgrounds, panel layouts, front-facing A-poses or 3D render look."
  4. Location pictures: describe what is in them (left side, centre, right side, background), and say the light and colour match it: "The colours and light match <Picture 3> exactly: neutral daylight, warm ochre walls, pale blue sky, no teal or green colour cast."
  5. Material or style pictures: "a waist-high wall made from the material of <Picture 3>: dark fractured rock with glowing orange cracks".
  6. Object pictures: give the real-world size and how it is held (section 5.4). Objects lifted from product shots tend to come out oversized.
  7. Real people in a scene: "Nobody poses and nobody looks at the camera" gives a candid result. Mention natural skin texture, sweat, dust, fabric wear.
  8. Each picture costs time and memory. Use only the pictures that add something.
  9. Enhance prompt off.

8.3 Common jobs

Job Key wording
Person into a new place (no location picture) <Subject 1> is the woman in <Picture 1>, same face and hair. She sits at a window table in a small Lisbon café...
Person into a pictured place ...<Picture 2> is the location. Place <Subject 1> at the table on the right, lit by the window light of <Picture 2>.
Two people together Define both subjects, then their positions: <Subject 1> on screen-left in a right-facing three-quarter view, <Subject 2> on screen-right facing her...
Outfit from one picture onto a person from another <Subject 1> is the man in <Picture 1>, same face and build. He wears the exact jacket from <Picture 2>: ... Use only the jacket from <Picture 2>.
Product into a lifestyle scene The bottle is the exact perfume bottle from <Picture 1>, same shape, cap and label text "...", about 12 cm tall, standing on a marble bathroom shelf...
Consistent character across images Reuse the same subject definition paragraph, word for word, in every prompt.

8.4 Pose or composition reference

To copy a pose from a picture without copying the person: "<Subject 1> takes exactly the pose of the person in <Picture 2>: ... Use only the pose from <Picture 2>, not their face, clothing or background." Also describe the pose in words.

8.5 Edit + references

When one picture is the base to change and others supply things to insert:

  • Base picture is Picture 1. Start with "Edit <Picture 1>: keep the exact location, camera angle, perspective and light of <Picture 1> - ."
  • Then "Add into this exact scene", and introduce each inserted item from its picture with position, size, facing and depth order.
  • End with how the new parts take the light of Picture 1.

9. No negative prompt: how to steer away from things

The model runs at CFG 1 with no negative prompt, so "negative" control happens only inside the positive prompt.

  1. Describe what should be there instead. Not "no clutter" but "a bare white desk with only a laptop and a mug". Not "not blurry" but "sharp focus on her eyes".
  2. Do not list unwanted objects in detail. Naming a thing can summon it.
  3. A short closing clause is fine for style drift and extras, because the text encoder is an LLM and understands it: "not illustration, not CGI render look", "no text overlays, no watermark", "no other people". Keep it to one short line at the end.
  4. For edits, use "keep" lists rather than "don't change" lists.

10. Problems and fixes

Problem Fix
Glossy, plastic "doll" face Longer, physical prompt; name a phone or film camera; "visible skin texture and pores, no retouching" (4.1)
Distant faces broken (teeth, eyes) Render at 16 MP; or move the person closer; give their figure size
Person looks pasted in Light direction on the face matches the scene light; cast and contact shadows; same colour temperature as the room
Wrong body direction Place the target in the frame, then state camera-relative facing (5.1)
Too many people Exact count; "no other people in the frame"; background people "small, seen from behind"
Reference face drifts or two faces merge Subject labels (8.2); describe the face in words too; fewer references
Character-sheet look leaks in (grey backdrop, A-pose, game render) "Use only their faces and outfits. Do not carry over..." (8.2.3)
Object from a reference is huge Real size in cm and a comparison with a body part
Colour cast (teal, green, orange) "neutral white balance, true-to-life colours, no colour grading", or "colours match exactly"
Text misspelled Double quotes, shorter text, larger size, one text item per location
Edit changed more than asked Longer, more concrete keep list; "Everything else stays exactly the same"; smaller edit
Image looks generic Replace adjectives with specific objects, materials and wear
Too dark / flat Name a light source with direction and a fill: "bright window light from screen-left, white wall bouncing soft fill on the right"

11. Final check (run before answering)

  • Mode is right, and an edit starts with "Edit <Picture 1>:".
  • Aspect ratio chosen from the request and main subject, not from added detail.
  • Every request from the user is visible in the prompt.
  • Exact count of main subjects; adults stated as adults.
  • Every subject has a place in the frame, a camera-relative facing, and a clear action.
  • Hands and contact are stated where they matter.
  • One named light source with direction; shadows agree with it.
  • Camera: lens, height, angle, focus.
  • Quoted text is exact, placed, and short.
  • References: each picture is labelled, described in words, and told what to take and what to leave.
  • Edits: concrete keep list, change described as a result, light and perspective matched.
  • No keyword spam; no long list of unwanted things.
  • Output in the format of section 0, with settings.

12. Examples

12.1 Text-to-image, realistic person

Mode: text-to-image · Settings: 3:4 · draft 4 MP, final 16 MP · Enhance off

A candid smartphone photograph of an adult woman in her late twenties sitting at a small round marble table outside a Paris café on a grey spring afternoon. She has light olive skin with a few freckles across her nose, dark brown shoulder-length wavy hair with loose flyaway strands, brown eyes, and no visible makeup. She wears an oversized camel wool coat over a white crew-neck T-shirt and thin gold hoop earrings. She sits facing the camera in a front three-quarter view, turned slightly toward screen-left, her right hand wrapped around a white espresso cup on the table and her left elbow resting next to a folded copy of a newspaper. She smiles with her lips closed, looking just past the camera. Behind her, slightly out of focus, wicker café chairs, a green-painted shop front with gold lettering, and two passers-by seen from behind on the wet pavement. Soft overcast daylight from the open sky above and to screen-left, gentle shadows under her chin and on the right side of her face. Shot on the iPhone 16 Pro main camera at eye level from across the table, visible skin texture and pores, natural phone HDR, no retouching, ordinary Instagram photo, not cinematic, not studio.

12.2 Text-to-image, poster with text

Mode: text-to-image · Settings: 2:3 · 16 MP · Enhance off

A minimalist vintage travel poster in flat screen-print style with a limited palette of deep teal, sand, coral and cream. The upper two thirds show a lone white lighthouse with a red lantern room standing on dark rocks at screen-right, a low coral sun setting over a calm teal sea at screen-left, and three small seagulls drawn as simple cream shapes in the sky. A thin cream border runs around the whole poster. At the top, the headline "CAPE SOLACE" is set in large cream condensed sans-serif capitals with wide letter spacing. In the bottom band of solid sand colour, the line "Where the coast road ends" sits centred in small teal italic serif, and below it the small text "EST. 1874" in coral capitals. Subtle paper grain and slight ink misregistration at the colour edges.

12.3 Edit, one picture

Mode: edit · Pictures: Picture 1 = the user's living room photo · Settings: same ratio as Picture 1 · Enhance off

Edit <Picture 1>: keep the exact room, camera angle, perspective and furniture of <Picture 1> - the grey fabric sofa against the back wall, the wooden coffee table, the tall bookshelf on screen-left and the window on screen-right. The light is now late golden hour: warm low sun enters through the window on screen-right and throws long orange rectangles of light across the floor and the front of the sofa, while the left side of the room falls into soft blue-grey shade. A sleeping orange tabby cat is curled up on the right sofa cushion, about the size of the cushion, lit by the sunbeam, with a soft shadow on the fabric beneath it. Everything else in <Picture 1> stays exactly the same.

12.4 Multi-reference composition

Mode: multi-reference · Pictures: Picture 1 = man (character sheet), Picture 2 = woman (photo), Picture 3 = market street · Settings: 3:2 · 16 MP · Enhance off

<Subject 1> is the adult man in <Picture 1>, a real human with exactly the same face: tanned skin, short black hair with a fade, a trimmed black beard and dark brown eyes. He wears a rust-orange bomber jacket over a black T-shirt, black cargo trousers and white sneakers.
<Subject 2> is the adult woman in <Picture 2>, with exactly the same face: pale skin, freckles, green eyes and copper-red hair in a loose low bun. She wears a cream knitted cardigan, a sage-green slip dress and brown leather ankle boots.
<Picture 1> is a character sheet. Use only his face, hair and outfit; do not carry over its grey background, panel layout, A-pose or 3D render look.
<Picture 3> is the location: a narrow evening street market with fruit stalls under striped red-and-white awnings on the left, strings of warm bulb lights overhead, and a tiled old building with blue shutters at the far end.

A candid real-life photograph taken in the street from <Picture 3> just after sunset. <Subject 1> stands on screen-left in a right-facing three-quarter view, holding a paper bag of oranges against his chest with his left arm and offering one orange to <Subject 2> with his right hand. <Subject 2> stands on screen-right in a left-facing three-quarter view, laughing, reaching with her right hand to take the orange; their fingers almost touch in the centre of the frame. A few shoppers in the background are small and seen from behind. Warm light from the bulbs overhead and the stall lamps falls on their faces from above and screen-left, with cool blue dusk sky light filling the shadows. The colours and light match <Picture 3>. Shot handheld at eye level with a 35mm lens, f/2, both of them sharp, the far end of the street softly blurred. Natural skin texture with pores, real fabric, true-to-life colour, nobody looks at the camera.

13. Workflow notes (ComfyUI)

These match the common ComfyUI setup for this model; other hosts may name things differently.

  • Sampler: 6 steps, CFG 1.0, scheduler simple. LCM-type sampler (s_noise 1.1 → 0.9) gives the most real skin; plain lcm at 4 steps or euler at 6 also work. Do not raise the steps.
  • Size: set width/height on the MiniMax H3 Image Conditioning node; both multiples of 32. Draft at 4 MP, final at 16 MP.
  • Reference slots: switch a Load Image slot on (Ctrl+B un-bypasses it), fill from slot 1 upward. Keep the VAE connected to the conditioning node so faces and rooms do not drift.
  • Enhance prompt (upsampler): a small LLM expands a one-line idea into a long caption and picks a ratio. Useful for text-to-image when you have no LLM at hand; turn it off for edits and references. If you used this guide, your prompt is already long, so leave it off.
  • Seed: fix the seed while tuning wording, so changes come from the prompt, not the noise.