Modelpublic

MLXBits/ltx-2.3-10eros-v1.5-dmd-hybrid-mlx-q8

sourceHugging Faceotherupdated 2mo agoView on Hugging Face

0likes61downloads

Model Card

LTX-2.3 10Eros v1.5 DMD Hybrid v2 — MLX int8 (q8)

Int8 (8-bit) MLX quantization of **TenStrip/LTX2.3-10Eros** v1.5 with the DMD hybrid v2 distillation (`LTX2.3_DMD_hybrid_v2.safetensors`) baked directly into the transformer, replacing the native rank-384 distilled LoRA.

Converted with mlx-forge and packaged for **ltx-2-mlx** — a pure-MLX port of LTX-2 for Apple Silicon.

[!WARNING] Not for all audiences. 10Eros is intended for adult use. By downloading and running this model you confirm you are of legal age in your jurisdiction and accept responsibility for the content you generate. Do not use it to produce illegal material or to depict real, identifiable people without consent.

What this is

This is a distilled-only package. The hybrid v2 distillation deltas were merged into the 10Eros v1.5 base at strength 1.0, producing a standalone distilled transformer. There is no dev transformer and no runtime LoRA fusion — the distillation is already in the weights.

Run it with `--distilled`. The two-stage (--two-stage, --two-stages-hq) and one-stage-dev (--one-stage) paths are not available here — they require a dev transformer, which this package intentionally omits.

Why this pairing

10Eros v1.5 is v1.4 with a linear ramp across the attn1/attn2 gates, reopening the explicit detail that v1.2 and Sulphur had while keeping v1.4's base-model-style prompting and structural stability. Upstream's own note is that because v1.5 is a hybrid of a model built for DMD and a model built for the LTX distilled LoRA, it goes best with a hybrid LoRA — and names LTX2.3_DMD_hybrid_v2 as the one that currently works nicely. That is exactly what is fused here.

What "hybrid v2" means

The base is the JoyAI Echo DMD extraction (LTX2.3_DMD_reshaped_r256, rank 256, reshaped from the LTX 384 1.1 distilled-LoRA shapes, audio branch sourced from the 384 1.1 distilled LoRA). Hybrid v2 then replaces the early blocks with the stock 384 1.1 distilled LoRA, so the early motion/composition direction comes from the distilled LoRA while the later blocks stay pure DMD. Per the metadata carried in the file we merged:

Blend field	Value
`blend_base`	`LTX2.3_DMD_reshaped_r256.safetensors`
`blend_source`	`ltx-2.3-22b-distilled-lora-384-1.1.safetensors`
`blend_mode`	`replace`
`blend_layers`	`0-31` (of 48 transformer blocks)
`blend_alpha`	`0.77` video · `audio_strength` `1.0`
`blend_scope`	`attn1, ff, attn2, modulation, audio_attn1`
`output_rank`	`256`

(The upstream card rounds this to "blocks 0-29 replaced with 384 at 0.8 strength video / 1.0 audio"; the values above are what the shipped file actually records.)

Versus the plain DMD extraction, v2 is described as giving cleaner delta directions and finer motion detail — particularly on repetitive explicit motion — with slightly stronger overall movement than standalone DMD, without letting the distilled LoRA redraw or reintroduce its base-model tendencies. Versus the native 384 distilled LoRA, the DMD side still avoids the resampling-to-base drift, the conditioning drop from the latent, and the look overwrites / extra detail added at upscale refine.

Expect the conditioning balance to shift relative to a pure-DMD package: upstream suggests stronger conditioning with a less aggressive upscale, or fewer steps with lighter preprocessing.

What's in here

File	Size	Role
`transformer-distilled-1.1.safetensors`	~19.2 GB	Distilled transformer with the DMD hybrid v2 deltas pre-fused, int8
`connector.safetensors`	~5.9 GB	Gemma → DiT embedding connectors
`spatial_upscaler_x1_5_v1_0.safetensors`	~1.0 GB	1.5× neural latent upscaler
`spatial_upscaler_x2_v1_1.safetensors`	~950 MB	2× neural latent upscaler (stage-2 refine)
`vae_encoder.safetensors` / `vae_decoder.safetensors`	~1.4 GB	Video VAE (8× temporal, 32× spatial)
`temporal_upscaler_x2_v1_0.safetensors`	~250 MB	2× temporal upscaler
`vocoder.safetensors`	~250 MB	BigVGAN v2 vocoder + BWE generator
`audio_vae.safetensors`	~106 MB	Audio VAE decoder

Quantization: int8, group size 64, applied only to nn.Linear inside transformer_blocks. AdaLN, projections, connectors, VAE and vocoder remain bf16 (MLX cannot quantize Conv layers).

Text encoder: Gemma 3 12B is not bundled — ltx-2-mlx loads it separately via mlx-lm.

No distilled LoRA files ship with this package: nothing would fuse them (there is no dev transformer, and --distilled doesn't need a distilled LoRA).

Usage

Requires ltx-2-mlx on Apple Silicon.

bash

# Text-to-video (distilled two-stage: half-res → upscale → full-res refine)
ltx-2-mlx generate \
  --model /path/to/ltx-2.3-10eros-v1.5-dmd-hybrid-mlx-q8 \
  --prompt "your prompt" \
  --distilled \
  -H 480 -W 704 -f 97 -o out.mp4

# Image-to-video — add --image
ltx-2-mlx generate \
  --model /path/to/ltx-2.3-10eros-v1.5-dmd-hybrid-mlx-q8 \
  --prompt "animate this" \
  --distilled --image photo.jpg -o out.mp4

Default distilled flow is the usual 8/4-step upscale schedule. Any Euler / LTX-compatible sampler works — no custom loading or sampling is needed. On <=32 GB Macs add --low-ram for block-streamed inference; q8 fits 16 GB with --low-ram.

Target sigmas

Upstream's suggested schedules for this distillation:

Pass	Sigmas
First pass	1.000, 0.955, 0.893, 0.812, 0.715, 0.603, 0.482, 0.241, 0.121, 0.0
First pass, higher quality	1.000, 0.968, 0.926, 0.875, 0.812, 0.741, 0.661, 0.574, 0.482, 0.241, 0.121, 0.000
Upscale	0.92, 0.725, 0.421875, 0.0
Upscale, less aggressive	0.77, 0.55, 0.29, 0.0

Prompting follows 10Eros: strict, descriptive, directive. LTX contributes little of its own — first frame, every motion, every audio cue has to be commanded.

Conversion provenance

1.LTX2.3_DMD_hybrid_v2 was merged into the 10Eros v1.5 bf16 base at strength 1.0. All 1660 LoRA pairs matched a base weight (none skipped), and every layer's .alpha equals its own rank (256 / 128 / 32), so the per-layer scale alpha ÷ rank is exactly 1.0 throughout. Both the video and audio branches were fused, since hybrid v2 carries a full-strength audio blend.

bash

   uv run ~/mlx-forge/scripts/merge_lora.py
   # LoRA:   LTX2.3_DMD_hybrid_v2.safetensors
   # Base:   10Eros_v1.5_bf16.safetensors
   # → 10Eros-v1.5-distilled-dmd-hybrid-bf16.safetensors

1.Distilled variant + all shared components + upscalers, quantized to int8:

bash

   uv run mlx-forge convert ltx-2.3 --variant distilled-1.1 \
     --checkpoint /path/to/10Eros-v1.5-distilled-dmd-hybrid-bf16.safetensors \
     --quantize --bits 8 \
     --spatial-upscaler x2 x1.5 --temporal-upscaler x2 \
     --output models/ltx-2.3-10eros-v1.5-dmd-hybrid-mlx-q8

mlx-forge validate ltx-2.3 <dir> passes with 0 warnings.

No dev variant is added — this package is distilled-only by design.

Credits & license

—DMD distillation & hybrid blend: TenStrip — LTX2.3_DMD_Lora · 10S-Comfy-nodes
—Hybridization prior art: maximsobolev275/LTX2.3-DMD-Distill-Lora-Hybrids
—Finetune: TenStrip/LTX2.3-10Eros
—Base model: Lightricks/LTX-2.3
—Conversion: mlx-forge · Runtime: ltx-2-mlx

Distributed under the LTX-2 license. The 10Eros finetune's own terms also apply — review the upstream model card before redistribution.