elismasilva/ltx2.3_image_custom_blocks

LTX 2 Image custom modular blocks

Custom Modular Diffusers blocks that extend LTX 2 Image with image-to-image, plus a unified AutoBlocks that folds text-to-image and img2img into a single pipeline. The workflow is chosen automatically from which inputs you pass.

prompt             -> text2image
prompt + image     -> image2image        (optional strength)

Loading & running

import torch
from diffusers import ModularPipeline

pipe = ModularPipeline.from_pretrained(
    "elismasilva/ltx2.3_image_custom_blocks",
    trust_remote_code=True,
)
pipe.load_components(
    names=["text_encoder", "tokenizer", "connectors", "transformer", "vae", "scheduler"],
    pretrained_model_name_or_path="elismasilva/ltx2.3-image-base",
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

image = pipe(
    prompt="a cinematic close-up portrait with crisp realistic detail",
    height=704,
    width=1280,
    num_inference_steps=28,
    guidance_scale=3.0,
    output="images",
)[0]
image.save("ltx2_image_t2i.png")

image-to-image

Pass an image and optional strength to the same pipe:

import torch
from diffusers import ModularPipeline
from diffusers.utils import load_image

pipe = ModularPipeline.from_pretrained(
    "elismasilva/ltx2.3_image_custom_blocks",
    trust_remote_code=True,
)
pipe.load_components(
    names=["text_encoder", "tokenizer", "connectors", "transformer", "vae", "scheduler"],
    pretrained_model_name_or_path="elismasilva/ltx2.3-image-base",
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

init_image = load_image("input.png")

result = pipe(
    prompt="preserve the scene while restoring crisp natural detail",
    image=init_image,
    height=704,
    width=1280,
    num_inference_steps=28,
    guidance_scale=3.0,
    strength=0.20,
    output="images",
)[0]
result.save("ltx2_image_i2i.png")

Blocks

The unified pipeline is composed of six modular blocks:

1. LTX2ImageTextEncoderStep encodes prompts into packed per-layer Gemma hidden states.
2. LTX2ImageConnectorStep adapts those hidden states with the image text connectors.
3. LTX2ImageVaeEncoderStep optionally encodes an input image for img2img.
4. LTX2ImagePrepareLatentsStep prepares one-frame image latents and FlowMatch timesteps.
5. LTX2ImageDenoiseStep runs the LTX 2 Image denoising loop.
6. LTX2ImageDecodeStep decodes latents into images or returns latent output.

Components

The blocks expect the following components from elismasilva/ltx2.3-image-base:

• text_encoder
• tokenizer
• connectors
• transformer
• vae
• scheduler

Diffusers merge compatibility

The image transformer reuses LTX2 attention, modulation and rotary helpers; those shared mathematical implementations were unchanged by the reviewed 2026-10-10 merge (7d047b180 versus be07836d9). The image transformer and image-only upstream pipelines were also unchanged. No video keyframe/DFR logic is added to this single-frame path. CPU contracts cover the reused helper signatures and loading the local block graph; image math and defaults remain intact. Shared loader/revision updates remain in Diffusers. New dependency versions require separate CUDA/performance approval.

Notes

The pipeline supports prompt embeds, custom sigmas or timesteps, guidance_scale, guidance_rescale, PAG inputs (pag_scale and pag_applied_layers), and VAE decode controls (decode_timestep and decode_noise_scale). For img2img, input_noise_sigma, input_sharpen, phase_cutoff, phase_transition_width, and phase_pad_factor can be used when you need more control over how the input image is converted into latents.