MiniMax H3 Reference Image Consistency: Same Character in Every Shot
The short answer: lock your character down with a reference image before you generate a single shot, then feed that same reference into every subsequent generation through the asset library and ComfyUI workflow so the model keeps the face, outfit, and build stable. On the MiniMax H3 导演台 (墨川导演台), consistency isn't a prompt trick — it's a pipeline habit: build the character in the asset library, generate a clean reference sheet, reuse it across first frames and tail frames, and let the cloud ComfyUI workflow carry continuity from shot to shot. Everything below is the practical version of that workflow.
---
Why does character consistency break in the first place?
It breaks because text prompts describe a person in words, and words are loose. "A woman in her thirties with short black hair" can render as a thousand different faces across a thousand generations.
AI video models don't have memory between separate generations by default. Each run starts fresh. If your only input is a text prompt, the model re-imagines the character every time — different jawline, different eye shape, slightly different age. In a single clip nobody notices. Across eight shots of a short drama, the audience absolutely notices.
The fix is to stop describing the character and start *showing* the character. That's what a reference image does: it converts a fuzzy text description into a concrete visual anchor the model can match against. On minimaxh3, reference-image support is built into the workflow, which is exactly why consistency is achievable here rather than a coin flip.
There are three layers to get right, and they stack:
- Visual anchor — a clean reference image of the character.
- Asset management — storing that reference so every task pulls the same file.
- Workflow continuity — using first frames, tail frames, and reference images together so shots connect.
Miss any one of these and consistency drifts.
---
How do I build a reliable character reference image?
Start in AI 制图 (AI image generation) and produce a clean, front-facing, evenly lit portrait — then save it to the asset library so it becomes the single source of truth for that character.
Here's the step-by-step:
- Open AI 制图 and generate a base portrait. Use text-to-image with a tight, specific prompt. Keep it neutral: front-facing, plain background, soft even lighting, neutral expression. You want a reference, not a hero shot.
- Refine with image-to-image if needed. If the first result is close but the face isn't quite right, feed it back through image-to-image and nudge it. Iterate until the face reads clearly at small size.
- Generate a small reference set, not one image. Produce a few angles — front, three-quarter, and a profile if you can get one. A single angle limits how the model handles shots where the character turns.
- Upload to the 资产库 (asset library). This is the critical step. The asset library is where you manage characters, scenes, props, and reference videos. Upload your character reference images there so they're reusable across every project and every generation task.
- Name it consistently. Use a clear character name so you don't grab the wrong file at 2 a.m. when you're batch-generating.
A few prompt habits that help the reference hold up:
- Describe fixed, unchangeable traits explicitly: face shape, hair length and color, distinguishing marks.
- Describe wardrobe as a locked costume, not a mood. "Navy wool coat, single breast" beats "winter outfit."
- Avoid words that invite variation — "beautiful," "stunning," "cinematic lighting" — in the reference prompt. Save those for the shot prompts.
The reference image is your character's ID card. Make it boring and precise.
---
How do I keep the same character across multiple shots?
Reuse the same reference image in every generation task and pair it with first-frame and tail-frame control so each shot starts and ends on-model.
The practical loop inside a project:
- Create the project in 项目库 (project library). Projects organize your generation tasks and versions, so all shots for one episode live together. This matters — you want the same reference applied across a batch, not re-uploaded per shot.
- Attach the character reference to each shot. When you generate a shot, pull the character reference from the asset library. The ComfyUI workflow on the cloud side uses that reference for consistency, matching the character rather than reinventing them.
- Generate first frames in AI 制图. For each shot, generate a still that establishes the opening composition — with the character reference applied. This still becomes the first frame.
- Generate tail frames where the shot needs to hand off. If shot A ends and shot B begins in the same space, generate a tail frame for A that visually matches B's opening. This is how 分镜衔接 (shot-to-shot continuity) actually works in practice.
- Batch generate the video. Use batch generation in the project to run the whole sequence with the same reference and frame settings, rather than one-off runs that drift.
- Review in AI 日志 (AI logs). Watch the generation log as tasks run so you catch a bad batch early instead of discovering it after twenty renders.
The mental model: reference image controls who the character is, first frame controls where the shot starts, tail frame controls where it ends. Consistency is the product of all three agreeing.
One more habit that saves hours: keep wardrobe changes deliberate. If your character changes costume between scenes, generate a *new* reference set for that costume and store it as a variant in the asset library. Don't let the model improvise a wardrobe change mid-scene.
---
Where does the cloud setup fit into all this?
Cloud deployment is what makes batch, reference-driven generation practical — you connect the director desk to your own GPU server over SSH and run the ComfyUI workflow remotely.
The 导演台 supports three run modes: local deployment on your own GPU machine, cloud deployment to a GPU server you rent, and API calls. For reference-heavy, batch-heavy work, cloud is usually the smoother path because the ComfyUI workflow handles reference-image consistency, shot continuity, and batch queuing on the server side.
If you're testing on 仙宫云, the flow is:
- Register at 仙宫云 and complete verification — note that the private image requires registering through the invitation link and entering invitation code B1H9M6.
- Send your 仙宫云 account ID to customer service (WeChat: ymcandai) to get the test account opened.
- Once the test account is active, you'll see the director desk image — deploy an instance with it.
- On the running instance, click SSH on the right side to get the host, port, and password.
- Enter those into the 云部署 (cloud deployment) page: SSH Host, SSH Port, SSH password.
- Click 验证 SSH (verify SSH). Once it passes, click 启动云端 ComfyUI (start cloud ComfyUI) and wait for the green success indicator.
- Now write your script and test a generation.
The local-tunnel port isolation means multiple users don't collide on the same machine — useful if you're working with a team. For anything about server specs, pricing, or availability, check with the official site or customer service rather than guessing.
---
How do I write shot prompts that respect the reference?
Keep shot prompts about action, camera, and light — never re-describe the character's face, because that fights the reference image.
This is the most common mistake. People write a gorgeous three-paragraph character description in every shot prompt, and the model averages that text against the reference image, producing a face that's a blend of both. The reference should win. So:
- Do describe: what the character is doing, where the camera is, what the light is doing, what the mood is.
- Don't describe: face shape, eye color, hair length, age. Those live in the reference.
- Do keep the costume consistent with the reference set, or explicitly switch to a stored costume variant.
- Don't introduce new physical traits mid-project unless you're deliberately changing the character.
A workable shot prompt skeleton:
[Action the character performs] + [camera angle and movement] + [lighting and time of day] + [environment details] + [mood/atmosphere]
Everything about *who* the character is comes from the reference. Everything about *what's happening* comes from the prompt. Keep that division clean and your consistency problems mostly disappear.
---
FAQ
Do I need a separate reference image for every costume?
Yes — treat each costume as its own reference variant in the asset library. A single reference can't reliably hold two different outfits, and letting the model improvise wardrobe is how continuity breaks.
Can I use the same reference across different projects?
The asset library is built for reuse across projects, so yes — but keep character naming consistent so you always pull the correct file when batch generating.
What if the character still drifts between shots?
Check three things in order: (1) is the reference actually attached to every task, (2) are your shot prompts re-describing the face (which fights the reference), and (3) do your first frames and tail frames actually match at the handoff points.
---
Consistency isn't a magic setting — it's a discipline of locking the character once and reusing that lock everywhere. Build the reference, store it in the asset library, attach it to every task, and let first frames and tail frames carry the continuity. That's the whole game with minimaxh3导演台.
Start at www.ymcdirector.com, add customer service on WeChat ymcandai for setup questions, and if you prefer shopping on Taobao, search 「杨墨川AI」.