Guidance too low with the embedder active. Raise guidance to 3.5-4 or add stronger light and color words.

Trends

.ai

les meilleurs modèles d'IA open source

Flex.1 Alpha by Ostris: the open, fine-tunable Flux cousin

Q: Dev LoRA causes artifacts

Partial compatibility. Lower the LoRA to 0.5 or use a LoRA trained on Flex.

Sujet : Modèles

Par Captain

Publié le 2026-09-30

Flex.1 Alpha is an 8B Apache-licensed model derived from Flux Schnell with a guidance embedder. Learn how its guidance differs, the settings that work, prompting, and why it is a favorite base for training.

Cette page n'est pas encore traduite ; elle est affichée en anglais.

Aperçu

Flex.1 Alpha by Ostris is an 8-billion-parameter rectified-flow model built from

FLUX.1

Schnell, with the distillation removed and a trainable guidance embedder added. It is released under Apache 2.0, which makes it one of the few Flux-class models that can be fine-tuned and used commercially without restriction.

Prompting is like Flux: natural sentences, no

negative prompt

by default. Because the guidance embedder is separate, you can bypass it and use true classifier-free guidance when you want a negative prompt.

Référence

Nom	Type	Rôle
Prompt	text	Natural sentences as with Flux.
Guidance	3-4	3.5 default with the embedder. Set to 1 and use CFG 3-5 if you bypass it.
Steps	25-50	It likes a few more steps than Dev; 30 is a good baseline.
Resolution	1 MP	1024x1024 and standard ratios.
LoRAs	Flex / Dev	Flex LoRAs are ideal; many Dev LoRAs work at reduced strength.

Aucune installation requise – exécutez les modèles IA dans le cloud

BitVector Prism est la façon la plus simple de commencer : choisissez un modèle, tapez un prompt et générez dans une application web épurée, sans rien configurer. BitVector est aussi disponible sur Discord et sur le web (SpyGlass).

Essayer BitVector

Pas à pas

Prompt in sentences and set guidance 3.5, steps 30.
Compare with the same seed at 50 steps; fine detail improves more than on Dev.
If you need to exclude something, bypass the guidance embedder and use CFG 4 with a negative prompt.
Add a Flex-trained LoRA at 0.8-1.0.
For training, use it as a base: its open license and un-distilled weights are the reason it exists.

Exemples

Illustration

A flat vector-style illustration of a lighthouse keeper's cottage at night, warm windows, stars, limited palette of navy, mustard and cream.

Photograph

A photograph of a mechanic's hands adjusting a bicycle derailleur, grease, soft workshop light, 50mm, shallow focus.

Creative idea: CFG with a negative

A busy farmers market in watercolor, loose washes, lots of white paper. Negative: photo, 3d render, text

Guidance embedder bypassed, CFG 4

Astuces

Flex keeps more of Schnell's bright, graphic look than Dev; it suits illustration and design.
Text rendering is decent but below Dev; keep words short.
Lower guidance (2.5) gives surprisingly natural photographs.
It is smaller than Dev and runs faster on the same hardware.

Dépannage

Output is washed out

Pourquoi cela arrive

Guidance too low with the embedder active.

Comment le corriger

Raise guidance to 3.5-4 or add stronger light and color words.

Dev LoRA causes artifacts

Pourquoi cela arrive

Partial compatibility.

Comment le corriger

Lower the LoRA to 0.5 or use a LoRA trained on Flex.

Aucune installation requise – exécutez les modèles IA dans le cloud

PirateDiffusion est réservé à Telegram et conçu pour les pros : des milliers de modèles, LoRA et workflows pilotés par commandes de chat, avec génération illimitée à prix fixe.