BREAKING NEWS
LATEST POSTS
-
Flex 1 Alpha – a pre-trained base 8 billion parameter rectified flow transformer
https://huggingface.co/ostris/Flex.1-alpha
Flex.1 started as the FLUX.1-schnell-training-adapter to make training LoRAs on FLUX.1-schnell possible.
-
Generative Detail Enhancement for Physically Based Materials
https://arxiv.org/html/2502.13994v1
https://arxiv.org/pdf/2502.13994
A tool for enhancing the detail of physically based materials using an off-the-shelf diffusion model and inverse rendering.
-
Camera Metadata Toolkit (camdkit) for Virtual Production
https://github.com/SMPTE/ris-osvp-metadata-camdkit
Today
camdkitsupports mapping (or importing, if you will) of metadata from five popular digital cinema cameras into a canonical form; it also supports a mapping of the metadata defined in the F4 protocol used by tracking system components from Mo-Sys.
-
OpenTrackIO – free and open-source protocol designed to improve interoperability in Virtual Production
OpenTrackIO defines the schema of JSON samples that contain a wide range of metadata about the device, its transform(s), associated camera and lens. The full schema is given below and can be downloaded here.
-
Martin Gent – Comparing current video AI models
https://www.linkedin.com/posts/martingent_imagineapp-veo2-kling-activity-7298979787962806272-n0Sn
🔹 𝗩𝗲𝗼 2 – After the legendary prompt adherence of Veo 2 T2V, I have to say I2V is a little disappointing, especially when it comes to camera moves. You often get those Sora-like jump-cuts too which can be annoying.
🔹 𝗞𝗹𝗶𝗻𝗴 1.6 Pro – Still the one to beat for I2V, both for image quality and prompt adherence. It’s also a lot cheaper than Veo 2. Generations can be slow, but are usually worth the wait.
🔹 𝗥𝘂𝗻𝘄𝗮𝘆 Gen 3 – Useful for certain shots, but overdue an update. The worst performer here by some margin. Bring on Gen 4!
🔹 𝗟𝘂𝗺𝗮 Ray 2 – I love the energy and inventiveness Ray 2 brings, but those came with some image quality issues. I want to test more with this model though for sure.
-
RigAnything – Template-Free Autoregressive Rigging for Diverse 3D Assets
https://www.liuisabella.com/RigAnything
RigAnything was developed through a collaboration between UC San Diego, Adobe Research, and Hillbot Inc. It addresses one of 3D animation’s most persistent challenges: automatic rigging.
- Template-Free Autoregressive Rigging. A transformer-based model that sequentially generates skeletons without predefined templates, enabling automatic rigging across diverse 3D assets through probabilistic joint prediction and skinning weight assignment.
- Support Arbitrary Input Pose. Generates high-quality skeletons for shapes in any pose through online joint pose augmentation during training, eliminating the common rest-pose requirement of existing methods and enabling broader real-world applications.
- Fast Rigging Speed. Achieves 20x faster performance than existing template-based methods, completing rigging in under 2 seconds per shape.
-
Skywork SkyReels – All-in-one open source AI video creation based on Hynyuan
https://github.com/SkyworkAI/SkyReels-V1
All-in-one AI platform for video creation, including voiceover, lipsync, SFX, and editing. One click turn text to video & image to video. Turns idea into stunning video in minutes. Check Pricing Details. Start For Free. All-In-One Platform.
SkyReels-V1 is purpose-built for AI short video production based on Hynyuan. It achieves cinematic-grade micro-expression performances with 33 nuanced facial expressions and 400+ natural body movements that can be freely combined. The model integrates film-quality lighting aesthetics, generating visually stunning compositions and textures through text-to-video or image-to-video conversion – outperforming all existing open-source models across key metrics.
-
Shanhai based StepFun – Open source Step-Video-T2V
https://huggingface.co/stepfun-ai/stepvideo-t2v
The model generates videos up to 204 frames, using a high-compression Video-VAE (16×16 spatial, 8x temporal). It processes English and Chinese prompts via bilingual text encoders. A 3D full-attention DiT, trained with Flow Matching, denoises latent frames conditioned on text and timesteps. A video-based DPO further reduces artifacts, enhancing realism and smoothness.
FEATURED POSTS
-
Gamma correction
http://www.normankoren.com/makingfineprints1A.html#Gammabox
https://en.wikipedia.org/wiki/Gamma_correction
http://www.photoscientia.co.uk/Gamma.htm
https://www.w3.org/Graphics/Color/sRGB.html
http://www.eizoglobal.com/library/basics/lcd_display_gamma/index.html
https://forum.reallusion.com/PrintTopic308094.aspx
Basically, gamma is the relationship between the brightness of a pixel as it appears on the screen, and the numerical value of that pixel. Generally Gamma is just about defining relationships.
Three main types:
– Image Gamma encoded in images
– Display Gammas encoded in hardware and/or viewing time
– System or Viewing Gamma which is the net effect of all gammas when you look back at a final image. In theory this should flatten back to 1.0 gamma.
(more…)
