obann001
Delphine-Cassarino-Flux-Rank32_lora Available on FriendliAI
Dedicated Endpoints Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Learn more Container Run this model inference with full control and performance in your environment.
Learn more Model Details
Model Provider
obann001
Model Tree
Base
black-forest-labs/FLUX.1-dev
Supported Functionality
Dedicated Endpoints Container
FriendliAI Corp:
San Francisco, CA
Copyright © 2026 FriendliAI Corp. All rights reserved
Delphine-Cassarino-Flux-Rank32_lora API & Inference Endpoint | FriendliAI
Summary
This LoRA is the Delphine Flux checkpoint selected at step01766 from the local Comfy training chain (Nemeris -> Delphine -> Emma) executed on 2026-10-07.dcsrn and keeps the base behavior of FLUX.1-dev for composition while biasing face/style identity.
Checkpoint : flux_file_delphine_32network128_rank32_bf16-step01766.safetensorsSHA256 : 561e4a883438d9730c75fbda79c061ca942c019c5c4e565ed18ff186843adbf8Size : 306,433,880 bytesTrigger token : dcsrnBase model : black-forest-labs/FLUX.1-devTraining stack : ComfyUI FluxTrainer workflow (flux_lora_train_delphine_clean.json)Dataset path used by run : /raid1_II/Pictures/Avatars_et_Persos/Delphine_Cassarino
Training Details
Table with columns: Field, Value Field Value Network dim / alpha 32 / 128Optimizer Adafactor (relative_step=False, scale_parameter=False)Learning rate 2e-4Precision bf16 (fp8 base enabled)Max train steps target
Evaluation Notes
A/B smoke comparison exists for step00800 vs step01766.
Local visual proofs:
../framework/assets/delphine-step00800-smoke.png../framework/assets/delphine-step01766-smoke.pngPortal section: ../framework/portal-nemokid.html (Comparatif Delphine)
Uploaded sample copies in this HF repo:
samples/delphine-step00800-smoke.pngsamples/delphine-step01766-smoke.png
Run chronology source:
Character Sheet (Style Guide)
Identity
Adult female portrait subject, dark skin tone, curly/wavy hair bias.
Tokenized identity anchor: dcsrn.
Intended for stylized-realistic portrait outputs, not strict face-lock biometric reproduction.
Visual Style
Strong portrait framing (head/shoulders, chest-up, half-body).
Natural skin rendering, moderate texture detail, non-plastic finish.
Balanced contrast, no aggressive HDR or oversharpen artifacts.
Palette
Warm neutrals: brown, amber, soft gold, terracotta.
Supporting cools: muted teal or slate backgrounds.
Prefer clean background separation for subject readability.
Prompting Baseline
Positive core : dcsrn, portrait photo, natural skin texture, curly hair, cinematic soft light, detailed eyesOptional style extenders : editorial portrait, studio lighting, shallow depth of field, 35mm lookNegative guidance : deformed hands, extra fingers, blurry face, wax skin, overprocessed, text watermark
Intended Use
Character-consistent portrait generation in private/creative pipelines.
Visual exploration and prompt iteration for Delphine-themed outputs.
Comfy/Flux workflows where LoRA switching between checkpoints is needed.
Limitations
Trained on a narrow identity domain; may reduce diversity if over-weighted.
Not suitable for impersonation claims, biometric identity verification, or surveillance use.
Best results depend on prompt hygiene and moderate LoRA strength.
Recommended Runtime Notes
Start with moderate LoRA strength and then tune by prompt family.
Validate across multiple seeds before selecting production renders.
Keep A/B checkpoints available (step00800, step01766) for regression checks.
Safety and Ethics
Respect consent and portrait rights for any subject-like generation.
Do not use this model for deceptive, defamatory, or non-consensual content.
Review outputs for bias/artifacts before publication.
1750
Selected checkpoint step01766
Class token injection dcsrn via workflow class_tokens
training_output/watch_nemeris_then_flux.log
09:31
16:18
