Seedreamstyle-H3-6K Image Generation Style LoRA: 4GB+ VRAM, File List

Sample outputs are available on the model card.
At a Glance
| Item | Value |
|---|---|
| Repository | Nebsh/Seedreamstyle-H3-6K |
| Family guide | MiniMax-H3 guide (15 articles) |
| Publisher guide | MiniMax: models and licenses |
| Published | 2026-11-20 |
| License | minimax-h3-image-community-license-agreement |
| Formats | safetensors |
| Source type | Unverified (not confirmed by a primary source) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Note: This news has not been officially confirmed and is unconfirmed information. The following is based on the description in the Hugging Face model card and is reportedly an unofficial public release by an individual (Nebsh), rather than an official announcement from MiniMaxAI.
Nebsh/Seedreamstyle-H3-6K is a style LoRA for MiniMaxAI/MiniMax-H3-Image. By adding seedreamstyle. to the beginning of the text prompt, it is trained to output human scenes (portraits, conversations, various age groups, practical lighting, and environmental shots) in a Seedream-style cinematic photographic tone. The author states that this is an experiment to see if the atmosphere of the closed-source image model Seedream 5.5 can be brought to H3 Image.
Specifications
- Base model:
MiniMaxAI/MiniMax-H3-Image(LoRA in adapter format) - Distribution file: 6K is recommended, with the filename
Seedreamstyle_H3_Resume_6K.safetensors - Trigger:
seedreamstyle.at the beginning of the prompt (common to all versions) - Verified strength: 1.0 in ComfyUI’s LoRA Loader (Model Only)
- Checkpoint used for verification:
minimax_h3_image_pruned_int8_convrot.safetensors - Comparison settings: LCM, 6 steps, CFG 1, scheduler simple. KJ’s
MiniMaxH3PromptWeight(mode: text with empty image, coverage 0-25%) is also used in combination - Output: H3 images in the comparison gallery were reportedly generated at 8448 × 4736 (approx. 40MP)
- Training: The 2K checkpoint was trained on the original 43-dataset, and steps 2,500 to 6,000 were continued on an expanded dataset of 1,134 images (including the original 43). The step count is the total including the initial 2,000 steps
- License:
minimax-h3-image-community-license-agreement. Conditions such as commercial use availability cannot be read from the documentation
The training settings are as follows (from the model card):
| Item | Value |
|---|---|
| Final global steps | 6,000 |
| Continued training | 2,000 → 6,000 (additional 4,000 steps) |
| Continued training dataset | 1,134 image/caption pairs (including the original 43) |
| Training resolution | Approx. 1 megapixel, aspect ratio buckets |
| LoRA rank / alpha | 16 / 16 |
| Optimizer | Automagic3 |
| Learning rate | 0.00028 |
| EMA | 0.98 |
| Batch / accumulation | 1 / 1 |
The base model card states that MiniMax-H3-Image is a 33.1B parameter DiT supporting outputs of approximately 16 megapixels.
Performance and Quality
Benchmark figures are not provided in the documentation. The model card includes a visual comparison gallery of 20 scenes by the author. Each board consists of four images: “No LoRA / LoRA 2,000 steps / LoRA 6,000 steps / Seedream 5.5", with both LoRAs set to a strength of 1.0. The author clarifies that this comparison is intended for “visual appearance and prompt interpretation" rather than a detailed performance comparison with matched resolutions. H3 is generated at approximately 40MP, while Seedream 5.5 is generated at 2736 × 1536 (native 2K), resulting in differing conditions.
The author states a preference for Seedream 5.5’s balance of contrast, lighting, and graininess. The extent to which this LoRA approaches that is based on the publisher’s own subjective comparison, and third-party verification has not been confirmed.
As an additional note, video experiments are included. Although this LoRA was trained for H3 Image, an “bonus experiment" trying 6K with H3’s video generation is published, showing an A/B comparison (No LoRA vs 6K strength 1.0, approx. 10–15 seconds) for T2V and Ref2VA. Compatibility with video is not guaranteed.
Strengths and Use Cases
This LoRA is suited for applying cinematic photographic expressions across various scenes featuring people. According to the model card, it aims to achieve styles with light expression, contrast, and photograph-specific graininess when generating portraits, conversation scenes, depictions of people of various ages, practical lighting setups, and shots including surrounding environments.
The primary task is text-to-image generation. The style is applied by adding the trigger word seedreamstyle. to the beginning of the prompt. It is adjusted to approach the textures of the closed model “Seedream 5.5" while leveraging the high-resolution and detailed rendering capabilities of the base model MiniMaxAI/MiniMax-H3-Image.
Additionally, application examples to video generation (T2V and Ref2VA) are published as a bonus experiment by the author, alongside still image generation. However, video application is not an officially verified feature and is positioned as experimental.
How It Differs from Similar Models
As previously introduced on this site, the “MiniMax-H3-Image-Character-Swap" Image Editing Model: 4GB+ VRAM Required is a model specialized for image editing that merges an adapter for character swapping based on the same MiniMax-H3-Image base model, allowing characters to be swapped while maintaining backgrounds and poses. In contrast, this model is a style-trained LoRA that imparts a cinematic photographic atmosphere and texture from text input.
Additionally, “Pika-H3-Flex" Audio-Enabled Video Generation Model: Distribution File List is a video generation model adapter based on the MiniMax-H3 series for generating videos with synchronized stereo audio. This model is a LoRA aimed at adding styles to still images, and its use cases and approach differ significantly from models whose primary function is the simultaneous generation of audio and video.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 4GB (laptop iGPU / phone class) | Original precision | 0.0GB | 0.0GB |
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. This release is an adapter (LoRA etc.); the table shows what the base model MiniMaxAI/MiniMax-H3-Image needs. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
The publisher distributes this model as safetensors.
License — minimax-h3-image-community-license-agreement: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.
Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Our Own Measurements
Measurements We Did Not Take
We have not confirmed that this site meets the commercial-use terms of this model’s license (minimax-h3-image-community-license-agreement). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).
Distributed Files
Weight files published in Nebsh/Seedreamstyle-H3-6K, listed by this site from the Hugging Face API. Sizes are the actual file sizes.
| File | Size |
|---|---|
checkpoints/Seedreamstyle_H3_2K.safetensors |
149MB |
checkpoints/Seedreamstyle_H3_step0002500.safetensors |
149MB |
checkpoints/Seedreamstyle_H3_step0003000.safetensors |
149MB |
checkpoints/Seedreamstyle_H3_step0003500.safetensors |
149MB |
checkpoints/Seedreamstyle_H3_step0004000.safetensors |
149MB |
checkpoints/Seedreamstyle_H3_step0004500.safetensors |
149MB |
checkpoints/Seedreamstyle_H3_step0005000.safetensors |
149MB |
checkpoints/Seedreamstyle_H3_step0005500.safetensors |
149MB |
checkpoints/Seedreamstyle_H3_step0006000.safetensors |
149MB |
Seedreamstyle_H3_Resume_6K.safetensors |
149MB |
How to Get It
Files for this model can be downloaded from the Hugging Face repository Nebsh/Seedreamstyle-H3-6K. The distribution format is safetensors. No prior license agreement procedure (gated) is required for acquisition.
The recommended file is Seedreamstyle_H3_Resume_6K.safetensors, which has been trained for 6,000 steps. Checkpoint files ranging from 2,000 to 6,000 steps are also provided within the repository for experimenting with and comparing the gradual changes during training.
Operation has been tested with ComfyUI as a compatible tool. It is recommended to load the model using ComfyUI’s LoRA Loader (Model Only) node and set the application strength to 1.0.
Related Articles
- MiniMax-H3-Image-Character-Swap Image Editing Model: 4GB+ VRAM
- MiniMax-H3-Character-Swap-LoRA Character Swap LoRA Adapter: File List
- Pika-H3-Flex Audio-Visual Video Generation Model: File List
- FastH3-Omni-Comfy: 80GB+ VRAM, File List
What to Read Next
- Explore the same model family → MiniMax-H3 family overview (15 articles, 4 converted builds)
- Formats this model is available in → Safetensors format guide and models
- Learn about the publisher → MiniMax: models, licenses and articles
- Other models for the same task → Other image generation models
Sources
This article contains unverified information. We will append an update note once it is confirmed by a primary source.

