MOSS Image Kodachrome V2

---

license: apache-2.0 tags:

• lokr
• lora
• text-to-image
• style
• kodachrome
• film

---

Kodachrome LoKr for MOSS-Image V2

This is a LoKr adapter that gives MOSS-Image the look of Kodachrome slide film. It mimics the color response and the contrast curve of 1940s 4x5 Kodachrome sheet film, capturing its quality and its sharpness. It was trained on MOSS-Image-Base with Ostris AI Toolkit at 1280 batch 4 for 1000 steps at low noise timestep bias.

V2 vs. V1

The first version, MOSS-Image-Kodachrome, was trained with the Balanced preset for timesteps. I found that this made the LoKr leak some composition properties from the dataset. I retrained it with mostly the same parameters, changing only to Low Noise timestep bias. After 1000 steps, the fine detail of the 4x5 sheet film, its tonal curve and its color balance had been learned, without the composition bleed.

V2 is more subtle than V1: no composition bleed, it applies mostly the colors and the tonal curve, and it accentuates fine detail. It is not necessarily better than V1, only different.

V1 was trained for 750 steps, about 3.5 epochs. The style had transferred by then, but not yet the grain, so treat V1 as an intermediate checkpoint.

Usage

Load it with the standard LoRA loader in ComfyUI (LoKr support is in core) or with load_lora_weights in Diffusers. Trained and tested on MOSS-Image-Base; tested at strength 1.0 with CFG 3, 50 steps, and skip-layer guidance on layer 22. Describe the scene in the prompt as you normally would; the film look comes from the adapter.

Trigger words

Use Kodachrome film or Kodachrome. A phrase like shot on Kodachrome film works better; a bare trigger may bleed into the image (see the front of the diesel engine).

Why a factor 16 LoKr

A film look is a small amount of information: a palette, a tone curve, grain. A full-rank LoKr at factor 16 has enough capacity for that and produces a file about 5% the size of a rank 32 LoRA. In community Krea 2 and MOSS runs, LoKr also shows less subject bleed than LoRA, which matters for a style trained on photographs of people and places; the adapter should carry the film and not the dataset.

Dataset

551 scans of Kodachrome slides, high resolution, no JPEG compression artifacts. Subjects and settings are deliberately varied: portraits, street scenes, landscapes, interiors, vehicles, daylight and artificial light, several decades of emulsion.

Images were trained at native resolution with aspect-ratio buckets, capped at a 1280² pixel area (1.64 MP); scans larger than that were downscaled, smaller ones left as they are. The set bucketed into 130 resolution groups, 66 of them single images. Captions describe the scene content and carry the trigger word; the film look itself is not described in the captions.

Training

Ostris AI Toolkit, architecture moss_image (MOSS-Image-Base), text encoder frozen.

• Network: LoKr, full rank, factor 16
• Optimizer: AdamW 8-bit, learning rate 1e-4
• Batch size 4 (within buckets, so 213 steps per epoch; mean effective batch 2.6)
• Steps: 1000, about 4.7 epochs (V1: 750)
• EMA: on, decay 0.98
• Timestep sampling: flowmatch, shift, low noise bias (V1: balanced)
• Resolution: 1280 (area cap), bucketing on, no square crop
• Precision: bf16
• Checkpoints every 250 steps; the 1000 step checkpoint is published here

Limitations

Grain and halation are learned at 1280² area and weaken at output sizes well below 1 MP. Inherits MOSS-Image's known weak points (small text, hands). Tested on Base only; behavior on MOSS-Image-SFT is unverified.

License

The adapter is released under Apache-2.0, matching the base model. The training images are not included.

Download

Get the files from the Files & versions tab.