July 2026 Summaries

A detailed exploration of a 3D-to-AI filmmaking process, the text describes the creation of a short action film featuring a 1930s pickup truck racing a steam train, utilizing an open-source 3DREAL Strong v2 adapter to achieve precise control over elements like animation, camera work, and timing. The pipeline is broken down into phases, starting with a 3D blockout stage using Blender controlled by AI, followed by generating photorealistic keyframes through image models, and culminating in rendering the final video with synchronized audio using the 3DREAL LoRA. The process emphasizes the necessity of iteration, human validation, and the use of various technological tools to ensure accuracy and coherence, ultimately allowing the filmmakers to achieve complete control over the visual and auditory components of their creation. This method showcases the potential of AI-driven tools in achieving artistic precision in filmmaking, contrasting with current video models that lack such detailed control.

FASHN is pioneering the integration of AI in fashion photography and virtual try-on, addressing longstanding logistical challenges in the industry by enabling brands to generate on-model imagery at scale without traditional constraints. Founded by Dan Bochman and his co-founder, the company allows fashion brands to produce, iterate, and localize photography efficiently, offering consumers the ability to virtually try on clothes before purchase. FASHN utilizes fal's serverless infrastructure to handle high-volume image processing, a critical feature for brands requiring consistent quality across large collections. While virtual try-on is particularly challenging due to the need for precise garment representation, FASHN has developed unique solutions, including building an open-source implementation for pixel-perfect editing. The company also navigates differing needs between brands, who prioritize accuracy, and consumers, who value speed, by ensuring their technology meets both demands without compromising on quality. Despite advancements, a significant challenge remains: accurately determining garment fit, which requires collaboration between brands and consumers. As the industry moves towards 3D-native design workflows, the potential for AI to bridge the gap between visual representation and actual fit grows, positioning FASHN as a leader in this evolving landscape.

The team at @fal has achieved a significant speedup in rendering Ideogram V4 images, reducing the time from 2.75 seconds to 0.44 seconds at a 1K resolution without compromising quality. This was accomplished through a series of innovations, including running the diffusion transformer in a 4-bit floating point format (FP4) and employing epilogue fusion to optimize memory usage during matrix multiplications. The process also involved quantization-aware distillation (QAD) to maintain image quality despite the 4-bit quantization, addressing issues like color desaturation that arose from early attempts. They further optimized the model by collapsing the traditional classifier-free guidance (CFG) into one forward pass and implementing timestep distillation, which significantly reduced the number of denoising steps required. These techniques collectively resulted in a model that achieves the same image quality as the full bf16 model but with a fraction of the computational cost, ultimately making the image generation process 6.3 times faster.

At Team fal, significant improvements were achieved in the throughput of the Qwen 3.6 model for Ideogram V4's prompt expander, aiming to enhance user interactivity in text-to-image generation. By leveraging DSpark on SGLang, they managed to increase throughput by 16 times, primarily through the adoption of a 35B MoE model that balanced inference speed and image performance. This model was fine-tuned using PEFT for efficiency and stability. Despite initial challenges with speculative decoding and acceptance rates, the team incorporated DFlash, a diffusion-based model, and later integrated DSpark to further boost performance. The DSpark architecture involved using diffusion-based block predictors combined with Markovian heads to improve acceptance rates, which required modifications to both vLLM and SGLang for efficient model serving. Through extensive experimentation, they achieved ~1000 tokens per second with DSpark and ultimately settled on a configuration that provided 830 tokens per second, maintaining interactivity and efficiency while reducing the prompt expansion time to under two seconds.