10Eros_v1.4 — INT8 Quantizations
INT8 quantizations of 10Eros_v1.4 (bf16), a fine-tune derived from LTX-2 (Lightricks). Two variants provided, both using row-wise scaling with heuristic layer exclusion.
Files
| File | Quantization | Size | Notes |
|---|---|---|---|
10Eros_v1.4_int8_row_heur.safetensors |
INT8, row scaling, --heur |
~22.3GB | Baseline quality, faster to produce |
10Eros_v1.4_int8_row_convrot.safetensors |
INT8, row scaling, --heur + --convrot (Hadamard rotation) |
~22.3GB | Higher quality target — ConvRot improves quantization accuracy at the cost of significantly longer conversion time |
Both load correctly with standard ComfyUI INT8-aware loaders — no special custom nodes required.
Details
- Source:
10Eros_v1.4_bf16.safetensors(~40GB) - Tool:
convert-to-quant(ctq), by SilverOxides - Format: ComfyUI-compatible (
--comfy_quant), quant metadata embedded - Tensors: 5,947 original → 1,774 weights quantized, 304 skipped (heuristically excluded as poor quantization candidates) → 8,887 final tensor count (both variants)
Which One Should I Use?
int8_row_heur: Good default. Clean output, no known issues.int8_row_convrot: ConvRot's Hadamard rotation typically produces measurably better quantization fidelity, especially on layers with high dynamic range. If you have the VRAM/patience to reproduce it yourself, worth trying first; both are provided here so you can compare directly.
Early testing found no meaningful quality gap between the two on standard SFW/NSFW test generations — pick based on what performs best for your specific use case.
Usage
Drop either file into ComfyUI/models/diffusion_models/, refresh node definitions, load via a standard ComfyUI diffusion model loader that supports --comfy_quant metadata (e.g. Load Diffusion Model (Quantized)).
License
This is a derivative of LTX-2 and inherits the LTX-2 Community License Agreement. Any use is subject to that license's terms — see link above.
Notes
Early INT8 conversions for LTX-based models — no LTX-specific preset yet exists in convert-to-quant; --heur (heuristic layer skipping) was used in place of a dedicated preset. ConvRot variant required an 80GB VRAM environment (RunPod A100 SXM) to complete due to memory demands on the largest embedding layers — not reproducible on typical consumer GPUs (tested: OOMs on 16GB).