Diffusae 2

Redraw your footage with Stable Diffusion and text prompts, right in your editing app.

After Effects Premiere Pro DaVinci Resolve Nuke Product page →

1. Description

Diffusae 2 brings generative AI into your host app. It takes each frame of the layer and redraws it from your text prompts (image to image), so you can restyle footage, turn sketches into images or edit a shot with a plain instruction.

It comes with more than ten models you download from inside the effect: Stable Diffusion 1.5 and 2.1, a depth-aware Stable Diffusion 2, fast LCM and Turbo versions, an SDXL-based model, a temporal consistency model, Instruct Pix2Pix, Illusion Diffusion, ControlNet Scribble and Canny, and two FLUX models. You can also add your own checkpoint models from Civitai.

Blend several prompts with their own weights, keep things out with a negative prompt, and choose how far the result moves away from the original with Strength. A fixed Seed keeps the result repeatable.

Requirements

Mac: Apple Silicon or Intel. Uses the GPU (Metal) when available.
Windows 64-bit: uses an NVIDIA GPU (CUDA) when available, otherwise the CPU, which is very slow for diffusion.
An internet connection to download models. Each model is about 5.5 to 7.5 GB; the FLUX models are about 17 GB (FLUX Mini) and 24 GB (FLUX.2 Klein).

2. Installation

2a. Baskl installer (all host apps)

Every Baskl plugin installs through the Baskl installer. It sets up AI Machine, the local engine the plugins run on, and puts Diffusae 2 in the right place for After Effects, Premiere Pro and OFX hosts (DaVinci Resolve, Nuke). Close your host apps before running it.

macOS

Click Download on baskl.ai, unzip Baskl.Installer.zip and open Mac/Baskl Installer.app. You need an Apple Silicon Mac; the installer stops on Intel Macs.
If macOS asks, confirm you want to open an app downloaded from the internet, then enter your Mac password so the installer can write to the system plug-in folders.
Click Let's Get Started, select Diffusae 2 in the plugin list and click Install. The first install also downloads AI Machine, so give it a few minutes.
When you see Installation Complete, restart your host app. The effect is in the Baskl category.

Windows

Click Download on baskl.ai, unzip Baskl.Installer.zip and run Windows\Baskl Installer.exe.
Click Yes when Windows asks to allow the app to make changes. The installer needs administrator rights.
Click Let's Get Started, select Diffusae 2 in the plugin list and click Install. On NVIDIA cards the installer picks the matching CUDA build for you. The first install also downloads AI Machine, so give it a few minutes.
When you see Installation Complete, restart your host app. The effect is in the Baskl category.

The installer does not download any diffusion model. After installing, click Download Models in the effect and download the models you want. Downloads run in their own window with a progress bar and a Cancel button, so your host app stays usable.
Coming from an older Diffusae 2 install? On the first render, models from the previous install (including custom Civitai models) are moved over automatically, so you do not download them again.
FLUX.2 Klein needs an up-to-date Baskl installation. If it fails to load, update with the latest Baskl installer.

Activating your license

After Effects / Premiere Pro: in After Effects, click Register at the top of the effect in Effect Controls (next to Reset), paste your license key and click Activate. In Premiere Pro, click the small icon button at the top right of the effect in Effect Controls to open the same dialog. The license is saved for the whole computer, so registering in one of the two apps covers the other.
Resolve / Nuke: click Register at the top of the effect's controls, paste your license key and confirm. Once activated, the button reads License Info.
Where to find your key: if you bought on baskl.ai, it is in your purchase email and under Account → My Licenses, with a Copy button. If you bought on aescripts + aeplugins, activate through the aescripts + aeplugins manager app, or paste the key from your aescripts account into the same Register dialog.
Without a license, Diffusae 2 runs as a trial: everything works, but renders carry a watermark until you register. Activation needs an internet connection, and each key works on a limited number of computers (shown under My Licenses).

3. Parameters

Model

Model: Click Choose Model to pick the model. The control shows the current model name; the default is SD-Turbo. Models that are not downloaded yet are marked Install first. Custom Civitai models appear at the bottom of the list.

Temporal Consistency SD 1.5 (Realistic Vision) with LCM, for fast and steadier frames.

Fast Stable Diffusion 1.5 (1.5 + LCM LoRA) The original Stable Diffusion model, sped up with LCM.

Fast Stable Diffusion 2.1 (SD-Turbo) A much faster distilled version of Stable Diffusion 2.1. The default.

Fast Stable Diffusion XL (SSD-1B + LCM LoRA) A distilled SDXL model, sped up with LCM.

Stable Diffusion 1.5 The original Stable Diffusion model; works well at smaller resolutions.

Stable Diffusion 2.1 Higher quality than 1.5 with the same memory needs.

Stable Diffusion 2 (Depth) Keeps the composition by estimating a depth map of the frame.

Instruct Pix2Pix Edits the frame from an instruction, like make it a winter scene.

Illusion Diffusion Uses the layer as a black and white layout that shapes the composition.

ControlNet (Scribble) Uses the layer as a rough sketch of the image.

ControlNet (Canny) Uses the layer as a Canny edge render of the image.

FLUX.2 Klein (4B) Distilled FLUX.2 model.

FLUX Mini A small distilled FLUX.1 model.
Download Models: Opens the model window: download, reinstall or delete the included models, open the models folder, and download custom Checkpoint models from Civitai by id, name or URL (an optional Civitai API key can be entered for login-only models).

Prompts

Prompt: Click Edit Prompts to describe what you want. Add one or more prompts, each with a Strength: prompts are blended by their relative strengths (two prompts at 100 and 50 weigh 2 to 1). FLUX models join the prompts and ignore the strengths.
Negative Prompt: Click Edit Negative Prompt to describe what you want to avoid. Only used when Guidance Scale is above 1; the fast LCM and Turbo models, Instruct Pix2Pix, the ControlNet models, Illusion Diffusion and FLUX ignore it.

Generation

Strength: How far the result moves away from the original frame, 0 to 100% (default 70%). Low keeps the footage, high follows the prompt. Not used by Instruct Pix2Pix, the ControlNet models or Illusion Diffusion.
Quality: Number of diffusion steps: more steps, more detail, slower render. The step count depends on the model (for example 14, 20 and 30 steps on the standard Stable Diffusion models, far fewer on the fast LCM, Turbo and FLUX models).

Normal Fastest. The default.

High More steps.

Very High Most steps, slowest.

Custom Use the step count set in Custom Steps.
Custom Steps: Step count when Quality is Custom, 1 to 49 (default 20).
Seed (-1 = random): Starting noise for the generation (default 42). Keep a fixed value for repeatable, steadier results; -1 picks a new random seed on every render.

ControlNet

Used by Illusion Diffusion.

ControlNet Strength: How strongly the layout image shapes the result, 0 to 200% (default 150%).
Guidance Start: Point in the generation, 0 to 100%, where the layout starts guiding it (default 0%).
Guidance End: Point in the generation, 0 to 100%, where the layout stops guiding it (default 100%). Lower it to give the model more freedom in the final details.

Advanced

Guidance Scale: How strictly the model follows the prompt, 0 to 50 (default 7.5). Higher sticks closer to the prompt. The fast LCM and Turbo models ignore it.
Image Guidance Scale: Instruct Pix2Pix only: how closely the result stays to the original frame, 0 to 10 (default 1.35). Higher keeps more of the original.
Scheduler: The sampling method. Leave on Auto unless you know you want a specific one. FLUX models use their own and ignore this.

Auto (Recommended) LCM for the fast LCM and Turbo models and Illusion Diffusion, Euler for the others.

PNDM, DDPM, DDIM, Heun, Euler, LMS, LCM, DPM Multistep Force a specific scheduler.
Internal Render Size: Resolution the model works at. The result is scaled back to the layer size.

Auto (Recommended) The size the selected model works best at (for example 512 for SD-Turbo, 1024 for FLUX).

Same as Layer Size No resizing. Slow on large layers.

2048, 1536, 1024, 768, 512 A fixed working size.
Resize Method: How the frame is fitted to the Internal Render Size. Not used with Same as Layer Size.

Auto (Recommended) Fastest above 768, Balanced at 768 and below.

Fastest The longest side matches the render size.

Faster Between Fastest and Balanced.

Balanced Halfway between Fastest and Highest Resolution.

Higher Resolution Between Balanced and Highest Resolution.

Highest Resolution The shortest side matches the render size. Slowest.