aleph-diffusion-adapters — the production artifacts of the diffusion aleph line

This is the curated adapters repo. The full research record (16 experiment packages, raw ledgers, 2-seed program) is geolip-aleph-diffusion; the framework that loads everything here is amoe-lora (pip install amoe-lora[diffusion]); the article is Part 3-D.

Every file is an amoe.diffusion.anchor safetensors (canonical blocks.{site}.{param} layout, full provenance in the metadata, content-hash verified). Nothing here modifies trunk weights: adapters attach to a frozen model, toggle off bit-exactly, and detach with verification.

The artifacts

file	trunk	what it is	evidence status
sd15/relay_all16_s0/.s1	SD1.5 core (eps)	the certified 16-site relay stack — grounding +0.1809→+0.2172 over zero-shot, beats matched LoRA which traded grounding away	exp006, candidate s0 + s1 replication
sd15/mb3_s0/.s1	SD1.5 core (eps)	the multiband stack: 3 band experts/site on the sigma axis; band lesions SURGICAL (own-band damage 50–200× cross-band); the step-gated controller moved image-space grounding +0.089 over frozen	exp008 2-seed + exp010 battery
sd15-lune-flow/relay_all16_s0	SD15-Lune flow	the first certified diffusion relay (−0.85% paired val, LoRA control WORSE than frozen)	exp001, candidate s0 (s1 replicated vs frozen)
sd15-lune-flow/blob_mb3_s0/.s1	SD15-Lune flow	the conditioning-law stack: blob-supervised HIGH band at λ≈1 (−8.3% on the role-aligned foreground gauge, common gauge in bound)	exp013, 2-seed
anima/aleph_relay_epoch4	Anima 2B DiT (flow)	28-block bf16 relay trained through the diffusion-pipe fork (native LoRA control DEGRADED while this improved)	exp004, candidate s0 · NC (CircleStone NC + NVIDIA OML — derived checkpoint)

Honest scope: these are research-grade adapters from a four-day 2-seed campaign, certified on their gauges (paired flow/eps-MSE, CLIP round-trip grounding, band lesions, the role-aligned foreground gauge) — not aesthetics-tuned style adapters. What they buy is measured and cited row by row; read the linked packages before deploying.

Quickstart

# pip install amoe-lora[diffusion]
import torch, amoe.diffusion as ad
from diffusers import StableDiffusionPipeline
from huggingface_hub import hf_hub_download

pipe = StableDiffusionPipeline.from_pretrained(
    "stable-diffusion-v1-5/stable-diffusion-v1-5",
    torch_dtype=torch.float32).to("cuda")

h = ad.attach(pipe.unet, hf_hub_download(
    "AbstractPhil/aleph-diffusion-adapters", "sd15/mb3_s0.safetensors"))
img = ad.sample(pipe, h, "a lighthouse at dusk", seed=7)   # step-gated

with h.lesion_band(2):        # generate WITHOUT the HIGH/structure band
    img2 = ad.sample(pipe, h, "a lighthouse at dusk", seed=7)
base = h.detach()             # verified bit-exact, or it raises

The relay stacks attach the same way (no windows needed). The sd15-lune-flow/ artifacts belong on the Lune flow trunk (ckpt-2500) with objective="flow" sampling — the metadata records the trunk id; ad.attach warns on substrate mismatch.

ComfyUI

Planned as comfyui-amoe (next release): loader / attach / band-toggle / detach nodes consuming exactly these files, with step gating installed as a pre-forward hook so any stock KSampler becomes step-gated. The node asserts each anchor's site count + width signature from the metadata before patching — a silent enumeration mismatch loads plausibly and corrupts quietly, so it ships only behind an image-parity check.

Training your own

Single GPU / DDP: amoe.diffusion.train (five lines — see the amoe-lora card). Multi-GPU production: the diffusion-pipe fork with aleph_relay = true (relay + multiband modes; saves land in this repo's format).

Licenses

MIT for this repo's own content and the SD1.5-family adapters. The anima/ adapter is a DERIVED checkpoint of NC-licensed weights (CircleStone NC + NVIDIA Open Model License) — non-commercial use only.