Two weeks! That’s how long it’s been since I pushed an update. The longest since launch. And I’ve been grinding!
Huge focus on H3 speed across GPUs, + Flux3 & other goodies!
Maestro v3.0 Updates:
• Faster generation and better memory management. Optimized kernels deliver roughly 1.5–2× speedups in our matched benchmarks, with major improvements on lower-VRAM machines.
• Smarter Auto-Tune and longer clip recommendations. Settings and recommended durations account for GPU, VRAM, RAM, resolution and workload—including tested 720p Fused generation on 8GB GPUs.
• Improved Director planning and prompt enhancement. Better story progression, character consistency and continuity, fewer unnecessary planning failures, and optional lyric guidance for music videos.
• New image models. Flux 3 Klein Distilled and Base support generation and editing with up to 10 reference images. Qwen Image 2.1 Turbo is also available.
• Visual Prompt Helper. Draw composition boxes, select reference-image regions and mark objects for removal to build clearer image-editing instructions.
• Expanded H3 capabilities. 2× VAE upscaling, LightVAE and additional sparse-attention choices, including experimental Veda.
• Fish Audio Drama 3. Expressive dialogue with emotions, whispers, laughter, pauses and voice cloning from up to three reference recordings.
• More capable local LLM options. Underdog Saluki 27B joins the available writers, with optional experimental MTP acceleration for compatible models. It’s a highly optimized 8B model based on & benchmarks around qwen3.8 27B but works on smaller GPUs.
• Better Director editing and regeneration. Load individual shots into Studio, retain prompt drafts through failed reruns, replace takes and rejoin the production.
• A cleaner everyday interface. Models organized by media type and mode, simpler names and defaults, expandable Director clip groups, and video durations visible before playback.
Two weeks! That’s how long it’s been since I pushed an update. The longest since launch. And I’ve been grinding!
Huge focus on H3 speed across GPUs, + Flux3 & other goodies!
Maestro v3.0 Updates:
• Faster generation and better memory management. Optimized kernels deliver roughly 1.5–2× speedups in our matched benchmarks, with major improvements on lower-VRAM machines.
• Smarter Auto-Tune and longer clip recommendations. Settings and recommended durations account for GPU, VRAM, RAM, resolution and workload—including tested 720p Fused generation on 8GB GPUs.
• Improved Director planning and prompt enhancement. Better story progression, character consistency and continuity, fewer unnecessary planning failures, and optional lyric guidance for music videos.
• New image models. Flux 3 Klein Distilled and Base support generation and editing with up to 10 reference images. Qwen Image 2.1 Turbo is also available.
• Visual Prompt Helper. Draw composition boxes, select reference-image regions and mark objects for removal to build clearer image-editing instructions.
• Expanded H3 capabilities. 2× VAE upscaling, LightVAE and additional sparse-attention choices, including experimental Veda.
• Fish Audio Drama 3. Expressive dialogue with emotions, whispers, laughter, pauses and voice cloning from up to three reference recordings.
• More capable local LLM options. Underdog Saluki 27B joins the available writers, with optional experimental MTP acceleration for compatible models. It’s a highly optimized 8B model based on & benchmarks around qwen3.8 27B but works on smaller GPUs.
• Better Director editing and regeneration. Load individual shots into Studio, retain prompt drafts through failed reruns, replace takes and rejoin the production.
• A cleaner everyday interface. Models organized by media type and mode, simpler names and defaults, expandable Director clip groups, and video durations visible before playback.