Davidyich/otp3d-style-loras

OTP 3D Style LoRAs

The Ostris AI Toolkit catalog contains 133 trained LoRA checkpoints (21.14 GB) from 13 OTP 3D style training runs. The separately maintained Civitai archive below is additional to these counts.

The original model folders and checkpoint filenames are preserved. S, M, and L identify composition complexity, not image resolution. Separate Klein 9B L/M runs remain in separate folders.

Models

Folder	Base model / run	Style	Trigger	Rank	Files	Final export step
ideogram-lm-1	Ideogram 4, combined L+M, RTX 5090 run	LM	otpbaselm	32	21	6000
ideogram-l1	Ideogram 4, RTX 5090 run	L	otpbasel	32	12	3000
ideogram-m1	Ideogram 4, RTX 5090 run	M	otpbasem	32	9	3000
ideogram-s1	Ideogram 4, RTX 5090 run	S	otpbases	32	9	3000
flux-klein9b-m1	FLUX.2 Klein Base 9B	M	otpbasem	16	8	3000
flux2-klein9b-m2	FLUX.2 Klein Base 9B, A40 run	M	otpbasem	16	11	5000
flux2-dev-l	FLUX.2 Dev	L	otpbasel	16	5	2500
flux2-klein9b-l	FLUX.2 Klein Base 9B, run 1	L	otpbasel	16	8	3000
flux2-klein9b-l2	FLUX.2 Klein Base 9B, A40 run	L	otpbasel	16	9	3000
z-base-l	Z-Image Base	L	otpbasel	16	7	3000
z-turbo-l1	Z-Image Turbo, A40 run	L	otpbasel	32	13	5000
z-turbo-m1	Z-Image Turbo, A40 run	M	otpbasem	32	9	5000
z-turbo-s1	Z-Image Turbo, A40 run	S	otpbases	32	12	5000

Choosing and downloading a checkpoint

Each folder contains a README with checkpoint steps and download links. The filename without a numeric step suffix is the final export for that run. Numbered files are intermediate checkpoints; their training steps are recorded in the manifest.

Download one checkpoint using the Hugging Face CLI:

hf download Davidyich/otp3d-style-loras \
  flux-klein9b-m1/otpbasem_flux2_klein9b_style_v1.safetensors \
  --local-dir ./otp3d-models

Load the LoRA with the matching base model shown above and include its exact trigger: otpbasel, otpbasem, otpbases, or otpbaselm for the combined Ideogram L+M run. These files contain LoRA adapters rather than full base-model weights.

Archive contents

• Published .safetensors checkpoint files from 13 source model folders. The Ideogram L+M run includes steps 1000–6000; earlier checkpoints are intentionally excluded.
• manifest.json: file sizes, SHA-256 checksums, training steps, epochs, tensor types, ranks, and model mapping.
• SHA256SUMS: checkpoint checksums in the standard checksum-file format.

Optimizer state, sample images, sample archives, and local training data are outside this checkpoint publication.

Training provenance

Exported by Ostris AI Toolkit: version 0.13.25 for FLUX/Z-Image runs and 0.13.34 for Ideogram 4. Model identifiers and trigger words were checked against the training configurations. Export steps and LoRA ranks were checked against the checkpoint metadata and tensor headers.

Ideogram 4 L/M/S used separate JSON-caption datasets (36/40/32 images), 512 px training resolution, 3000 steps, rank/alpha 32/32, learning rate 0.0001, and Differential Guidance scale 3. Use an Ideogram 4 compatible loader; its unconditional training adapter is not a style checkpoint.

Ideogram L+M combines 36 L and 40 M images with JSON captions (76 pairs), trigger otpbaselm, 6000 steps, 512 px, rank/alpha 32/32, learning rate 0.0001, batch/accumulation 1/1, AdamW8Bit, and Differential Guidance scale 3. The final step is verified in SafeTensors metadata; model quality has not been ranked against the separate L/M runs.

Duplicate source files

The following original filenames contain identical bytes and are retained to preserve the archive:

• flux2-klein9b-l/otpbasel_flux2_klein9b_style_v1_000002250 (1).safetensors, flux2-klein9b-l/otpbasel_flux2_klein9b_style_v1_000002250.safetensors

Recovered Civitai archive

Additional recovered checkpoints and original settings exports from Civitai are preserved in a separate namespace. The root manifest and SHA256SUMS index only the Ostris catalog above; Civitai maintains its own inventory and checksums.

Archive and provenance, Civitai manifest, checksums.