Linarix-v2

Usage

Requires diffusers >= 0.38.0 — earlier versions have a trust_remote_code RCE (advisory). For production, pin a commit hash with revision= so the remote code cannot change under you.

Install

pip install -U "diffusers>=0.38.0" transformers accelerate safetensors torchvision scipy
pip install "flash-linear-attention @ git+https://github.com/fla-org/flash-linear-attention.git@3c4c54ae7397d37130d7101edd0f4eb596af896d"

FLA is required, not optional: the GDN-2 mixers and Block AttnRes need this exact build. Image attention runs on PyTorch SDPA by design — flash-attn is not used and installing it changes nothing. The STORK scheduler is bundled with this repo no separate install.

Generate

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "Akrao9/Linarix-v2",
    custom_pipeline="pipeline_boomer",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
).to("cuda")

image = pipe("a lighthouse on a rocky cliff above crashing waves at golden hour")[0]
image.save("output.png")

Every sampler setting defaults to the value this model was tuned and showcased at (STORK4, 20 steps, cfg_scale=4.0, flow_shift=5.0), so passing nothing reproduces the grid above. Override only what you want to change:

image = pipe(
    "a quiet cobblestone street in an old European town, evening",
    seed=42,          # None (default) draws a fresh seed each call
    cfg_scale=3.0,    # lower = softer, less saturated; raise for prompt adherence
    steps=32,         # more steps buy little past the default
)[0]

The transformer weights come from this repo. The DC-AE VAE and Qwen/Qwen3.5-4B text encoder are fetched from their upstream repos on first use — run hf auth login first if this repo is gated for you.

Batched inference

Pass a list of prompts to generate a batch in one call:

images = pipe([
    "a lighthouse above crashing waves",
    "a red fox in fresh snow",
    "a steam locomotive on a stone bridge",
])
images[0].save("a.png")

Two things matter for throughput. VAE slicing is on by default, decoding one image at a time so batched-decode peak memory stays flat; toggle with pipe.disable_vae_slicing(), or add pipe.enable_vae_tiling() for large images on low VRAM (both work before the VAE is lazily loaded). And keep components resident on the GPU when benchmarking — the default offload_text_encoder=True moves the text encoder to CPU after each call, which is the right trade for VRAM but adds per-call transfer overhead that dominates small batches.

Samples

All 1024px, EMA weights, STORK4 / 20 steps, derivative_order=1, substeps=14, cfg_scale=4.0, cfg_rescale=0.5, cfg_interval=[0.1, 0.9], and flow_shift=5.0. Prompts (left→right, top→bottom):

1. a snow-covered mountain village at blue hour, warm windows glowing
2. a desert canyon at sunset, layered red rock walls
3. a wooden pier stretching into a misty lake at dawn
4. a stone castle on a green hilltop under drifting clouds
5. an elderly fisherman mending nets on a harbour wall
6. a ginger cat asleep on a sunlit windowsill
7. a vintage motorcycle parked on a rain-slicked city street at night
8. a field of sunflowers under a bright summer sky
9. a narrow canal in Venice with weathered facades, late afternoon
10. a steaming bowl of ramen on a dark wooden table
11. a great horned owl perched on a bare branch at dusk
12. a glass greenhouse full of ferns, soft diffused light

Licensing

The weights are released under a custom license; the terms are not finalised, so no open-source license is asserted here. Treat the tag as "all rights reserved pending the published terms" and ask before relying on these weights in anything you ship.

The runtime code bundled in this repo (pipeline_boomer.py, modeling_boomer_fla.py, qwen_connector.py) is covered by the same pending terms. STORKScheduler.py and STORK_constants/ are third-party (Tan et al. 2025, arXiv:2505.24210) and keep their own upstream license.

The text encoder (Qwen/Qwen3.5-4B) and the VAE (mit-han-lab/dc-ae-f32c32-sana-1.1-diffusers) are separate downloads under their own licenses; nothing here relicenses them.