Community texture graft for MiniMax H3

MiniMax H3 x Z-Image: Give Every H3 Video Film-Grade Texture

Some AI clips look great in motion yet feel a little plastic when you pause them: walls too clean, metal too smooth, skin like wax. MiniMax H3 x Z-Image, styled "MiniMax-H3 × Z-Image" by its creator, is a free community edition of the H3 audio-video model that tackles exactly that. It borrows the eye for fine texture from Z-Image, a still-image model famous for crisp surfaces, and hands it to H3. Faces, voices, speed and memory use stay the same; peeling paint, rust, wet pavement, fabric weave and pores simply come through richer. Here is what it changes, where it shines, and prompts that show it off.

  • Same H3 workflow
  • Richer surfaces
  • Stable multi-shot detail
  • GGUF + ComfyUI native
Make a texture-rich video →
Rich rusted engine room with flaking paint and glistening water droplets, showing the texture detail of MiniMax-H3 x Z-Image video

Uncensored Alternative Video Generator

Image to video

Upload a starting photo, describe the materials, light and motion you want, then generate a clip with sound.

A starting image is required for video.

The texture upgrade H3 fans were waiting for

H3 already moves beautifully and syncs sound with picture, but plenty of creators felt its surfaces came out a bit smooth. In August 2026 Hugging Face user joeygambino published MiniMax H3 x Z-Image, which the model card calls "the spatial detail graft". The idea is simple to explain: Z-Image, Alibaba Tongyi's 6B text-to-image model, bets hard on fine texture, so its habits were measured and copied onto the later layers of H3. Nothing is retrained and no new knowledge is added. The early layers stay untouched, because changing them caused grid-like artifacts on regular patterns.

For you, H3 x Z-Image means a drop-in file. The graft keeps H3's identity, voice, speed and VRAM footprint, and it plugs into the workflow you already use. H3 x Z-Image is an independent community build, not a release from MiniMax or Alibaba. File names such as zs05 and curve hint at the recipe: zs05 most likely means a spatial strength of 0.5, while curve is a per-layer blend that the author currently recommends. Neither name is fully explained on the card, so treat those readings as educated guesses.

Rust, rain and skin: where the extra detail shows

The author sums up the effect in one line: peeling paint peels more, rust sits deeper, water catches more light. The three H3 x Z-Image demo films on the model card, a greenhouse, a ship engine and a busy market, are each three shots long, so you can watch glass fogging, oil-stained steel and stacks of fabric hold their texture from cut to cut. With H3 x Z-Image, faces keep their identity and gain finer skin detail than stock H3 manages, which matters for close-ups, beauty shots and dialogue scenes.

Compared with the original model, the biggest win appears in long edits. When you chain clips, stock H3 tends to drift sharper each time, so the last shot looks over-processed next to the first. The author measured a high-frequency ratio of 0.99 after three extensions for MiniMax H3 x Z-Image, against 1.11 for the original; a value near 1.0 means detail stays steady instead of piling up. These are self-reported numbers with no independent benchmark yet, so the best test is your own: rerun an old prompt with the same seed and compare.

  • Weathered materials: flaking paint, corrosion, cracked plaster, worn leather.
  • Water and light: wet streets, condensation, droplets, neon reflections.
  • People: pores, freckles, make-up and the weave of linen, denim or silk.
Misty Victorian greenhouse with condensation on glass and wet leaves, a texture-rich scene for multi-shot AI video

Long multi-shot stories that look the same at the end

Many creators build 15 to 60 second sequences by linking several H3 clips. H3 x Z-Image works with the community H3 Multishot workflow for ComfyUI, which strings shots together seamlessly, and the author says no node changes are needed. Two H3 x Z-Image variants cover the classic needs: ref2va keeps a character's face and voice from a reference, and fl2va lands a clip on a chosen final frame. Both can be chained, so you can open on a wide establishing shot, cut to your lead in action, and end on an exact frame for a transition or reveal.

A simple way to write it: Shot 1 sets the place with material words and light direction; Shot 2 repeats the same character description while they act; Shot 3 describes the landing image for fl2va. Because detail in H3 x Z-Image no longer creeps sharper with each extension, you spend less time matching grades or softening later shots in post.

What creators are making with it

Anywhere viewers are close enough to notice materials, H3 x Z-Image earns its keep. These are the projects where the richer surfaces make the most visible difference.

Vertical short dramas

Interiors, costumes and close-ups read more like live action, while ref2va keeps the lead's face and voice from episode to episode.

Ads and product films

Brushed metal, leather grain, glass and food surfaces gain the tactile quality that sells a product in a three-second hook.

Atmosphere and world-building

Wastelands, steampunk workshops, rainy neon alleys and abandoned factories are made of grime and texture, so they benefit the most.

Realistic portraits and fashion

Skin texture, make-up and fabric folds look natural instead of airbrushed, a long-standing request for realistic skin from H3 users.

Prompt templates that bring out the texture

H3 x Z-Image only pays off when the prompt asks for something to render. Name concrete materials such as rust, linen, condensation or pores, say where the light comes from, and add a sound cue so H3 builds the matching audio. A popular first experiment is to rerun prompts you already know and look at the walls, metal and skin. Copy these and adapt them:

Material close-up
Close-up of a rusted ship engine, flaking red paint, oil stains, water droplets catching warm light, slow dolly in, ambient machine hum
Greenhouse mood
Inside a fogged Victorian greenhouse, condensation on glass panes, wet leaves, soft morning light through mist, a gardener walks past, birdsong
Market tracking shot
Busy night market, handwoven fabrics, steam from food stalls, neon reflections on wet pavement, handheld tracking shot, crowd chatter
Portrait texture
Portrait of an adult woman in her 30s, visible skin pores and freckles, linen shirt texture, golden hour window light, she turns and smiles, soft breathing
Rain and neon
Adult man with wet hair under a flickering neon sign, rain running down a leather jacket, puddles reflecting pink light, slow push-in, rain and distant traffic

GGUF or native: picking a file for 16, 24 or 32 GB

The graft comes in two download repos. The GGUF repo offers Q3mix, Q4_0, Q5_1 and Q8_0 quantizations in fl2va and ref2va flavours, with curve as the recommended build. The native repo holds ComfyUI-ready safetensors in bf16, fp8, mxfp8, int8, w4a8, w4a4 and nvfp4, and with about 24,000 downloads against roughly 5,000 for GGUF it is clearly the more popular way to run H3 x Z-Image. Put the H3 x Z-Image file in your usual H3 model folder, select it in the loader and keep the rest of your ComfyUI workflow as it is.

Your GPUSuggested GGUFNotes
16–24 GBQ4_0Lightest option the author lists for this range
24–32 GBQ5_1Balanced quality and memory
~32 GBQ8_0Closest to full precision
Recent NVIDIA cardsnative fp8 / nvfp4Pick the format your ComfyUI build supports

Choose ref2va to keep a person's look and voice, fl2va to end on a specific frame; curve is the current recommendation.

Skin, silk and swimwear: 18+ ideas and the uncensored question

H3 x Z-Image is not an uncensored fine-tune: it only rescales texture attention and neither adds nor removes content limits, and the card declares no NSFW focus. Still, adult creators like H3 x Z-Image, because sensual scenes live on exactly what it improves, namely skin, fabric and light. With original, fictional adult characters, try a white bikini on a sunlit beach with wet skin and sand texture, a lingerie studio set with satin folds, a steamy shower lit from the side, or silk pyjamas in a lamp-lit bedroom.

If you would rather not install anything, Spicy AI's image-to-video generator handles adult (18+) realistic clips with relaxed filters in the browser.

Beach (18+, fictional adult)
Adult model in a white bikini on a sunlit beach, wet skin glistening, sand texture, gentle waves, slow motion walk, ocean sound
Boudoir (18+, fictional adult)
Adult woman in silk lingerie, satin sheets, soft lamp light, fabric folds and skin detail, slow camera pan
Golden hour portrait of an adult woman with natural skin pores, freckles and linen fabric texture for realistic AI video

Creating texture-rich videos on Spicy AI

No GPU, no node graph: you can chase the tactile look of H3 x Z-Image online. The richer your prompt is in materials, the closer the result feels to real footage, which means fewer rerolls. Start from a strong still, for example a detailed keyframe, then let the generator animate it.

  1. 1Upload a reference image of your adult character, product or location.
  2. 2Write the shot with material words and light direction: rust, wet asphalt, linen, golden side light.
  3. 3Add the motion, camera move and sound you want, then generate.
  4. 4For a longer story, generate the next shot from your favourite frame and keep the same character description.

Turn a photo into a detailed clip with sound in a few clicks.

Open image to video

Self-hosting H3 x Z-Image vs generating online on Spicy AI

Running H3 x Z-Image locally gives you every node and file option; the online route trades that control for zero setup.

Local deployment (official weights)Spicy AI online
Content filtersYou configure your own setupRelaxed filters for adult content (18+)
Open weightsYes, public downloadNot needed — runs in the browser
Model you actually useThe graft on top of H3An online image-to-video model
Price to startYour own GPU (16 GB or more)Free tier available

FAQ

What is H3 x Z-Image in one sentence?

A community edition of MiniMax H3 released in August 2026 that copies Z-Image's fine-texture attention habits onto H3, so videos show richer surfaces while everything else stays the same.

Is it made by MiniMax or Alibaba?

No. It is a personal project by Hugging Face user joeygambino and has no official backing from either company.

Do I need a new workflow or any training?

No. The author says you can swap the H3 model file and keep your existing workflow, including the Multishot chaining setup.

Which file should I download?

Start with the curve build. Q4_0 suits 16–24 GB cards, Q5_1 suits 24–32 GB and Q8_0 needs about 32 GB; use ref2va to keep a character, fl2va to hit a final frame.

Does H3 x Z-Image work with my H3 LoRAs?

It stays inside the H3 family, so the usual acceleration and style LoRAs should load, but compatibility has not been verified for each one. Test on a short clip first.

Can I use it without installing anything?

The graft has no hosted service of its own. You can get a similar detailed look online with Spicy AI's image-to-video generator.

Sources

Browse all uncensored AI models →

Sign in

Welcome to Spicy AI

or