Comparison
MiniMax H3 Image Compared: How to Choose an Image Model and Migrate
A developer's checklist for picking an image model and switching without breaking your pipeline
На этой странице
What You Are Choosing
MiniMax H3 Image is a two-route image family: text-to-image and image-to-image. Both bill per image on the same official resolution tiers. Pick text-to-image when the request is prompt-only. Pick image-to-image when you must pass reference images (required image_urls, 1–16). The routes share resolution (default 2K), output_format (default jpeg), and seed, and they differ in the aspect_ratio default (1:1 vs auto).
Official Pricing
Official per-image resolution-tier prices, identical on both routes:
- 1K: $0.0129
- 2K: $0.0129
- 4K: $0.0393
On image-to-image, each reference image adds $0.0036. That surcharge is added to the tier price; it does not multiply the tier price. A 2K edit with four references costs $0.0129 + 4 × $0.0036 = $0.0273, not $0.0129 × 4. Text-to-image has no reference surcharge.
Because 1K and 2K share the $0.0129 rate, resolution as a cost lever mainly matters when you opt into 4K. Reference count is a separate, additive lever on the edit route.
Validation and cost guardrails after switching
Migration is not done when the first image comes back. It is done when you can prove the new pipeline behaves predictably and that its spend is bounded.
Validate with fixed seeds
The family's seed parameter makes a real regression test possible. The procedure:
- Pick a small suite of representative prompts — one per visual category your product actually ships.
- Run each prompt with an explicit
seed, explicitresolution, and explicitaspect_ratio, and store the outputs as your baseline. - Re-run the same configuration later and compare. Same seed plus same parameters should give you a comparable result; drift tells you something in your request changed.
- Change one variable at a time. Bump
resolutionfrom2Kto4Kand inspect; changeoutput_formatand inspect. Do not change three parameters and then argue about which one caused the difference. - Lock the validated configuration — a single exported constant object holding prompt, seed, resolution, aspect ratio, and format — and have services read from it rather than constructing requests inline.
Validate the reference path separately
Image-to-image needs its own suite, because positional prompt references are the most fragile part of the integration. Build a two- and a three-reference case, confirm the prompt grammar (Picture 1, Picture 2, …) addresses the right images, and confirm your ordering logic is deterministic. Also validate your pre-flight checks: JPG/PNG/WEBP, sides 128–8192 px, up to 20 MB each, and no more than 16 entries.
Guardrails against surprise spend
- Default to 2K in code, not just in the schema. Since 1K and 2K share a per-image rate, there is little reason to request 1K unless the consumer genuinely wants a small file. Pin
resolutionin your shared config so nobody accidentally inherits 4K. - Gate 4K behind an explicit flag. It is the tier with the higher per-image rate; it should require a deliberate opt-in from the calling code, not appear as a copy-paste default.
- Cap reference counts in your client. Enforce a maximum below the schema's 16 in your own validation layer, based on what your prompts actually need. Each reference adds $0.0036 on top of the tier price, so an accidental fan-out of references raises cost without adding value.
- Default to JPEG and treat PNG as an exception. Record the requested format in your output metadata so storage growth is explainable after the fact.
- Log per-request metadata. Prompt length, resolution, format, seed, reference count, and output bytes per request gives you everything needed to attribute spend and to spot a runaway job before the invoice does.
- Cap concurrency and batch sizes. Worker retries that duplicate generation requests are a common, self-inflicted cost source; make retries idempotent where possible and alert on retry rate.
Keep the loop closed
Treat the validation suite as a living artifact. Whenever you change a prompt template, a reference ordering rule, or a resolution default, re-run the seeded baseline and diff before shipping. That habit — fixed seed, explicit parameters, one change at a time — is what turns a migration from a hopeful cutover into a controlled one.
Pick the route that matches your workload, pin every parameter that affects cost, and the rest is ordinary plumbing.

Похожие статьи
Все статьи →-
Comparison
gpt-6.2-sol vs. comparable models: How developers should choose and migrate
Make a side-by-side decision across three dimensions: capability boundaries, pricing structure, and migration cost
-
Comparison
gpt-6.2-astra Side-by-Side Comparison and Selection Guide: How Developers Can Decide Whether It Is Worth Integrating
-
Comparison
Tripo H3.1 Side-by-Side Comparison and Selection Guide: How to Choose Among Three 3D Generation Paths and How to Save
A developer-oriented trade-off between pricing and capability, starting from input form, geometric refinement, and PBR/texture switches