- Seedreamstyle — MiniMax H3 Image LoRA · 6K
- Download and use
- A note about this experiment
- Available checkpoints
- Training
- Twenty-scene comparison gallery
- 01 Livre Du Soir
- 02 Classe Apres La Pluie
- 03 Attente Du Car
- 04 Repetition Ado
- 05 Quai De Fin D Ete
- 06 Atelier Coiffure
- 07 Laverie Neon
- 08 Rendez Vous Sous L Averse
- 09 Depart En Randonnee
- 10 Laboratoire En Fin De Service
- 11 Dispute A L Atelier
- 12 Anniversaire A Trois
- 13 Electricien Dans Les Combles
- 14 Pecheur Noir Et Blanc
- 15 Orage Sous La Veranda
- 16 Portrait Dans Le Bus
- 17 Quatre Generations A Table
- 18 Heure De Pointe
- 19 Fete Au Flash
- 20 L Aube Du Marche
- Video examples — T2V and character references
- License and attribution
- Download and use
Seedreamstyle — MiniMax H3 Image LoRA · 6K
A photographic style LoRA for MiniMax H3 Image, trained to bring a Seedream-like cinematic rendering to human scenes: portraits, conversations, varied ages, practical lighting and environmental shots. This is an independently trained H3 adapter; the Seedream 5.5 examples below are reference generations from Comfy Cloud.
All comparison images are AI-generated.
Download and use
- Recommended version: 6K. File:
Seedreamstyle_H3_Resume_6K.safetensors - Trigger:
seedreamstyle.at the start of your prompt. - Tested strength: 1.0 in ComfyUI's LoRA Loader (Model Only).
- Use with MiniMax H3 Image. The tested checkpoint was
minimax_h3_image_pruned_int8_convrot.safetensors. - Comparison recipe: LCM, 6 steps, CFG 1, scheduler simple.
- The comparisons use KJ's
MiniMaxH3PromptWeight, mode text with empty image, active over 0–25%.
Example:
seedreamstyle. An older couple talks on a glazed veranda during a storm. Waist-up two-shot, both faces visible. Cool storm daylight through rain-covered windows mixes with a small warm table lamp. Natural skin texture, restrained photographic grain and a quiet, attentive exchange.
A note about this experiment
I really like Seedream 5.5 — it's currently my favorite closed-source image model, especially the way it balances contrast, lighting and grain.
I was curious to see how much of that look I could bring to MiniMax H3 Image with a style LoRA, so I trained this little experiment.
Special thanks to @StefTranquillin for donating 47 images, and to @GrootDEX for providing around 100 images!
Available checkpoints
Earlier checkpoints are included so you can compare how the look develops and choose the version you prefer. Step counts are total training steps, including the initial 2,000 steps.
I recommend the 6K checkpoint. The other checkpoints are available if you would like to compare different stages of training.
| Total training steps | Download |
|---|---|
| 2,000 | Seedreamstyle_H3_2K.safetensors |
| 2,500 | Seedreamstyle_H3_step0002500.safetensors |
| 3,000 | Seedreamstyle_H3_step0003000.safetensors |
| 3,500 | Seedreamstyle_H3_step0003500.safetensors |
| 4,000 | Seedreamstyle_H3_step0004000.safetensors |
| 4,500 | Seedreamstyle_H3_step0004500.safetensors |
| 5,000 | Seedreamstyle_H3_step0005000.safetensors |
| 5,500 | Seedreamstyle_H3_step0005500.safetensors |
| 6,000 (recommended) | Seedreamstyle_H3_step0006000.safetensors |
The 2K checkpoint was trained on the original 43-image dataset. The 2,500–6,000-step checkpoints continue that training on the expanded 1,134-image dataset, including the original set. The trigger remains seedreamstyle. for every version.
The gallery below compares 2K and 6K at strength 1.0; the intermediate checkpoints are provided for experimentation and are not separately evaluated in that gallery. You only need to download the version you want to use.
Checkpoint sizes and SHA-256 hashes
The 6K file in checkpoints/ is identical to the main 6K download; you only need one copy.
Training
| Setting | Value |
|---|---|
| Final global step | 6,000 |
| Continuation | 2,000 → 6,000; 4,000 additional steps |
| Frozen continuation dataset | 1,134 image/caption pairs, including the original 43 |
| Target training resolution | Approximately 1 megapixel, aspect-ratio buckets |
| LoRA rank / alpha | 16 / 16 |
| Optimizer | Automagic3 |
| Learning rate | 0.00028 |
| EMA | 0.98 |
| Batch / accumulation | 1 / 1 |
Twenty-scene comparison gallery
Each board reads top left: Without lora; top right: LoRA 2,000 steps; bottom left: LoRA 6,000 steps; bottom right: Seedream 5.5. Both LoRAs use strength 1.0.
The comparison uses recorded descriptions, numeric seeds and inference settings. Seedream 5.5 uses the initial description with only the seedreamstyle. trigger removed. Exact per-version prompts and settings are included beneath each scene. Matching seed numbers across different models do not imply identical noise or compositions.
H3 images were generated at 8448 × 4736 (~40 MP); Seedream 5.5 references at 2736 × 1536 (native 2K). This repository's gallery images are reduced-size previews and comparison boards. The comparison covers appearance and prompt interpretation at these settings, rather than a matched-resolution detail benchmark.
Standalone gallery · Download the gallery ZIP · Comparison manifest
01 Livre Du Soir
Same original description and numeric seed.
02 Classe Apres La Pluie
Same original description and numeric seed.
03 Attente Du Car
Same original description and numeric seed.
04 Repetition Ado
Same original description and numeric seed.
05 Quai De Fin D Ete
Same original description and numeric seed.
06 Atelier Coiffure
Same original description and numeric seed.
07 Laverie Neon
Same original description and numeric seed.
08 Rendez Vous Sous L Averse
Same original description and numeric seed.
09 Depart En Randonnee
Same original description and numeric seed.
10 Laboratoire En Fin De Service
11 Dispute A L Atelier
Same original description and numeric seed.
12 Anniversaire A Trois
Same original description and numeric seed.
13 Electricien Dans Les Combles
Same original description and numeric seed.
14 Pecheur Noir Et Blanc
15 Orage Sous La Veranda
16 Portrait Dans Le Bus
Same original description and numeric seed.
17 Quatre Generations A Table
Same original description and numeric seed.
18 Heure De Pointe
Same original description and numeric seed.
19 Fete Au Flash
Same original description and numeric seed.
20 L Aube Du Marche
Same original description and numeric seed.
Video examples — T2V and character references
This LoRA was trained for MiniMax H3 Image. Video compatibility is a bonus experiment: I also tried the 6K checkpoint with H3 video. Below are matched A/B comparisons, plus A and B as separate watchable videos, each with its own original audio.
A = Without lora (0.0). B = Seedreamstyle 6K (1.0). Each pair uses the same prompt, seed, references where applicable, and inference settings. All characters and scenes are AI-generated fiction.
The side-by-side player uses B's audio by default. Play the individual A and B videos to compare their own voices, pauses and soundscapes. Generated timing and framing can differ even with the same seed.
Browse all video files · Video index
| Mode | Scene | A/B | A — Without lora | B — 6K |
|---|---|---|---|---|
| T2V | A bitter kitchen argument | Watch | Watch A | Watch B |
| Ref2VA | OpenCast conspirators | Watch | Watch A | Watch B |
| Ref2VA | A fictional town-hall debate | Watch | Watch A | Watch B |
| Ref2VA | An uneasy alliance | Watch | Watch A | Watch B |
A bitter kitchen argument · T2V · 10.1 s
Two spouses exchange bitter accusations in a kitchen before dawn.
A/B — left: Without lora; right: Seedreamstyle 6K at 1.0
A — Without lora · original audio
B — Seedreamstyle 6K at 1.0 · original audio
A/B file · A file · B file · Prompt and settings
OpenCast conspirators · Ref2VA · 10.1 s
Two eccentric fictional characters quietly plot together in a restaurant kitchen.
A/B — left: Without lora; right: Seedreamstyle 6K at 1.0
A — Without lora · original audio
B — Seedreamstyle 6K at 1.0 · original audio
A/B file · A file · B file · Prompt and settings · Character references
A fictional town-hall debate · Ref2VA · 10.1 s
A male candidate interrupts a female candidate with a sexist remark; she answers him.
A/B — left: Without lora; right: Seedreamstyle 6K at 1.0
A — Without lora · original audio
B — Seedreamstyle 6K at 1.0 · original audio
A/B file · A file · B file · Prompt and settings · Character references
An uneasy alliance · Ref2VA · 15.1 s
Two rival dealers consider an alliance in a backlit garage at sunset.
A/B — left: Without lora; right: Seedreamstyle 6K at 1.0
A — Without lora · original audio
B — Seedreamstyle 6K at 1.0 · original audio
A/B file · A file · B file · Prompt and settings · Character references
License and attribution
This trained adapter is a modification/derivative of MiniMax-H3-Image and is subject to the upstream MiniMax-H3-Image Community License Agreement, including its usage and territorial terms. See LICENSE and NOTICE. Base model: MiniMaxAI/MiniMax-H3-Image.
- Downloads last month
- 2,808



















