NeuroCognica

AI Model Licences

Last verified: 2026-07-31. Companion to THIRD_PARTY_NOTICES.md.

ChronoSophia does not redistribute AI model weights. This page tells you exactly which models the pipelines expect, where each one comes from, what its license permits, and which ones you must decide about before using Chronos for commercial work.

Licenses change. Every row links to the publisher's page — that page, not this one, is authoritative on the day you download.


1. Installed automatically (with your consent)

Model Used for License Commercial use Source
qwen2.5:7b-instruct (~4.7 GB) Turns your prompt into a validated scene plan Apache-2.0 Yes Qwen/Qwen2.5-7B-Instruct via Ollama

sophia:latest and sophia:v0 are local Ollama tags built from this model plus a Chronos prompt/persona file. They inherit Apache-2.0 from the base weights. The Chronos persona text itself is proprietary to NeuroCognica.

This is the only model the installer downloads by default. The setup script names it and its license before pulling, and will not proceed without your consent.

Optional, only if you tick the box (added 2026-07-27)

The installer offers a second checkbox, selected by default for the standard Museum Wall path: "Also download the AI image models (29 GB, Apache-2.0)". Untick it and setup skips those downloads; leave it selected and setup additionally fetches the FLUX.1 [schnell] set below. Leave it unticked and nothing here is downloaded — the recipes that need these models will say so plainly instead of rendering a blank canvas, and every other recipe works normally.

File Size License Source
flux1-schnell.safetensors 23.8 GB Apache-2.0 Comfy-Org/flux1-schnell
t5xxl_fp8_e4m3fn_scaled.safetensors 5.2 GB Apache-2.0 comfyanonymous/flux_text_encoders
clip_l.safetensors 0.25 GB MIT comfyanonymous/flux_text_encoders
ae.safetensors (the VAE) 0.34 GB Apache-2.0 Comfy-Org/Lumina_Image_2.0_Repackaged

On a machine with no NVIDIA GPU the installer fetches a different, smaller set instead — about 9 GB rather than 29:

File Size License Source
z_image_turbo_nvfp4.safetensors 4.5 GB Apache-2.0 Comfy-Org/z_image_turbo (upstream Tongyi-MAI/Z-Image-Turbo)
qwen_3_4b_fp8_mixed.safetensors 4.0 GB Apache-2.0 same repository
ae.safetensors 0.34 GB Apache-2.0 Comfy-Org repackage, hash-verified

Z-Image-Turbo is roughly half the size of FLUX and runs on a processor in minutes rather than hours. It is slower and less detailed than FLUX on a GPU — see SYSTEM_REQUIREMENTS.md. It is Apache-2.0, so commercial use of the model and of what you make with it is permitted, exactly as with schnell.

Why the VAE comes from a mirror. Black Forest Labs' own FLUX.1-schnell repository requires a HuggingFace account and a license click, so downloading from it would interrupt your install with a login. Comfy-Org republishes the identical file without that gate. We do not ask you to trust the mirror: the file is verified against SHA-256 afc8e28272cd15db3919bacdb6918ce9c1ed22e96cb12c4d5ed0fba823529e38 (335,304,388 bytes) and discarded if it does not match. The license is unchanged — Apache-2.0, commercial use permitted.

Every file in this set is resumable, so a dropped connection continues where it stopped rather than restarting 29 GB.


2. Optional / operator-supplied image models

Chronos's texture, concept, and painting pipelines call ComfyUI with the checkpoint filenames below. The optional installer image-model task may fetch the documented Flux or Z-Image sets in section 1 after the buyer's consent. Everything else is operator-supplied under the publisher's licence. If a checkpoint is absent, the pipeline that needs it reports the missing file — it will not silently substitute another model.

Checkpoint file Used by License Commercial use Source
dreamshaper_8.safetensors PBR/albedo material textures (DEFAULT_CHECKPOINT) CreativeML Open RAIL-M Yes, subject to the RAIL use restrictions Lykon/dreamshaper-8
z_image_turbo_bf16.safetensors Storyboard frames Apache-2.0 Yes Tongyi-MAI/Z-Image-Turbo
flux1-schnell.safetensors + ae.safetensors Default Canvas-on-Easel / Flux scene path (flux_schnell_scene) Apache-2.0 Yes The installer can fetch these for you — see the optional checkbox in section 1. Otherwise: black-forest-labs/FLUX.1-schnell (account required)
flux1-dev.safetensors + ae.safetensors Opt-in only via CHRONOS_FLUX_VARIANT=dev FLUX.1 [dev] Non-Commercial License Restricted — read section 3 black-forest-labs/FLUX.1-dev
clip_l.safetensors, t5xxl_fp8_e4m3fn_scaled.safetensors Text encoders for the FLUX / Z-Image workflows CLIP: MIT · T5: Apache-2.0 (verify on the repo you download from) Yes Published alongside the above workflows

Optional, only if you run scripts/download_ipadapter.py:

Model Used for License Commercial use Source
ip-adapter_sdxl.safetensors + image encoder Image-conditioned generation Apache-2.0 Yes h94/IP-Adapter
controlnet-union-sdxl-1.0 (promax) Structural control Apache-2.0 Yes xinsir/controlnet-union-sdxl-1.0

That script now prints these licenses and asks for confirmation before downloading anything.


3. FLUX defaults — commercial-safe schnell

Product default (sellable path): Canvas-on-Easel and ComfyUiClient::new_flux_scene load flux1-schnell.safetensors via flux_schnell_scene.json (Apache-2.0). That is the commercial-safe default.

Opt-in non-commercial path: set CHRONOS_FLUX_VARIANT=dev to select flux1-dev.safetensors / flux_dev_scene. That model is under the FLUX.1 [dev] Non-Commercial License.

For FLUX.1-dev only — two separate questions, two different answers:

The images you generate — commercially usable. BFL's license states that Outputs are not Derivatives of the model, and that you may use Output for any purpose, including commercial purposes, except as the license expressly prohibits.

Running the model itself — non-commercial only. The same license grants you rights to access, use, distribute, or create Derivatives of the model only for Non-Commercial Purposes. Using the model as part of a commercial activity requires a separate license from Black Forest Labs, which may carry a fee or revenue share.

What this means for you as a Chronos customer:

NeuroCognica is not affiliated with Black Forest Labs and cannot grant rights to their models. This section is a pointer to their terms, not legal advice.


4. Model outputs and your artifacts

Renders, .blend files, textures, and sealed artifacts Chronos produces are yours. NeuroCognica claims no ownership of your creative output and no license to it. Chronos does not transmit your prompts, models, or renders anywhere — everything runs locally (see docs/customer/SECURITY_AND_PROVENANCE.md).

The one constraint that travels with your output is whatever the model publisher's license says about its own outputs. For every model listed above except FLUX.1-dev, that constraint is either nil or the RAIL use-restriction list (no illegal use, no harassment, no deceptive impersonation, etc.). For FLUX.1-dev, see section 3.


5. Training weights (operator lane, not in the customer install)

Qwen/Qwen2.5-Coder-7B-Instruct (Apache-2.0, ~15 GB) is used only by the optional geometry fine-tuning pipeline. It is excluded from the default install by design and is never fetched on a customer machine.


Something wrong here? Report it. A misattributed license is a defect, and it gets corrected in the open. See docs/customer/SUPPORT.md.