base_model: black-forest-labs/FLUX.2-klein-base-9B tags: - text-to-image - flux - flux2-klein - lora - diffusers - template:sd-lora - art-history - aby-warburg - mnemosyne - pathosformel - visual-studies instance_prompt: mnemoatlas_pathos widget: - text: >- mnemoatlas_pathos, formula of grief, women in black with raised arms around a fallen body, fresco fragment and newspaper clipping, multiple black-and-white photographic plates on dark archival panel, loose grid layout, mid-century art-historical reproduction - text: >- mnemoatlas_pathos, formula of ecstatic frenzy, maenad with flying drapery and loosened hair carrying a dead child, ink drawing juxtaposed with political rally photograph, clipped reproductions on black panel, varying scale - text: >- mnemoatlas_pathos, formula of triumph, nude male figure standing on defeated enemy, Renaissance engraving beside modern parade photograph, dark archival layout, asymmetric grid - text: >- mnemoatlas_pathos, formula of the cosmic man, human body diagrammed as astral figure with arms open in cruciform posture, medieval manuscript beside early modern anatomical print, three photographic plates on black panel - text: >- mnemoatlas_pathos, formula of self-inflicted death, hand pressing dagger inward, Roman cameo beside daguerreotype, eighteen reproductions of varying scale dispersed across the dark panel library_name: diffusers pipeline_tag: text-to-image

Mnemosyne Atlas LoRA — Aby Warburg / Pathosformeln

Model description

A LoRA trained on the 63 panels of Aby Warburg's Bilderatlas Mnemosyne (1924–1929), fine-tuned on black-forest-labs/FLUX.2-klein-base-9B using ostris/ai-toolkit.

The Bilderatlas Mnemosyne is Warburg's unfinished masterwork: dark fabric panels on which hundreds of photographic reproductions are fixed with metal clips — antique reliefs, Renaissance engravings, newspaper photographs, advertising images — arranged without hierarchy between registers or epochs. The organizing principle is the Pathosformel (formula of pathos): the gesture charged with emotional intensity that migrates across centuries, from the Greek amphora to the press telegram, from Medea to the cabaret dancer.

This LoRA does not reproduce a style or a subject. It learns the compositional and conceptual grammar of the Atlas: black archival ground, irregular grid of B&W reproductions at varying scales, anti-hierarchical mixing of visual registers, specific gestural formulas recurring across time.

The dataset consists of 63 image–caption pairs, one per panel. Each caption (100–180 words, English) was built from Warburg's own source catalogue (~1090 sub-images across the 63 panels): trigger token + Pathosformel name + shared gestural grammar + 4–8 concrete citations mixing registers + compositional note. Text encoder: Qwen3-8B, native to FLUX.2-Klein.

Training: rank 16 / alpha 8 · 18 epochs · ~2270 steps · AdamW 8bit · batch 1 · bfloat16 · April 28, 2026.

Trigger word

Use mnemoatlas_pathos to activate the LoRA. Prompts without the trigger will produce standard FLUX.2-Klein output.

Recommended prompt structure

mnemoatlas_pathos, formula of [PATHOSFORMEL NAME],
[2–3 concrete gestural phrases],
multiple black-and-white photographic plates clipped on dark archival panel,
loose grid layout, varying scale, mid-century art-historical reproduction

Negative prompt:

text, watermark, logo, signature, bookshelves, library, color, blurry, low quality

Parameters: LoRA strength 0.75–0.90 · steps 20–28 · CFG 3.5–5.0

The seven Pathosformeln

Formula	Core gesture
Ecstasy / Fury	Loosened hair, raised arms, billowing mantle
Triumph	Standing figure over the defeated body
Cosmic vitality	Open cruciform body, astral-anatomical diagram
Terror / Mourning	Hands to face, conclamatio, lament
Heroic struggle	Torsioned body in combat
Stoic exit	Weapon turned against the self
Nachleben moderno	Ancient formula reappearing in press and advertising

Download model

Weights are available in Safetensors format (~158 MB).

Download them in the Files & versions tab.

Use it with the 🧨 diffusers library

from diffusers import AutoPipelineForText2Image
import torch

pipeline = AutoPipelineForText2Image.from_pretrained(
    "black-forest-labs/FLUX.2-klein-base-9B",
    torch_dtype=torch.bfloat16
).to("cuda")

pipeline.load_lora_weights(
    "[username]/mnemosyne-atlas-lora-flux2",
    weight_name="aby_klein_280426.safetensors"
)

image = pipeline(
    "mnemoatlas_pathos, formula of grief, women in black with raised arms "
    "around a fallen body, multiple black-and-white plates on dark archival panel",
    num_inference_steps=24,
    guidance_scale=4.0,
    generator=torch.Generator("cpu").manual_seed(42)
).images[0]

Dataset

The training set covers all 63 numbered panels of the Bilderatlas, including the introductory cartographic plate (panel A) and the Homo zodiacalis plate (panel B). Source data: Warburg Institute catalogue of the ~1090 reproductions distributed across the panels.