Release 4.7.0

SN34's earlier identity: Filed under SN34 from a source its on-chain identity named before. The identity has changed and no longer names this source, so the item is not listed as news of the subnet on SN34 now.

Release Notes Validator: prompt generation overhaul The validator's prompt generator (gas/generation/prompts/prompt_generator.py) has been re-architected from a rule-based motion / camera planner with hardcoded option lists into a VLM-grounded, LLM-composed pipeline. New architecture source image │ ▼ Qwen3-VL-4B-Instruct ───► SceneDescription (structured JSON: subjects, setting, lighting, dynamic_candidates, observed_motion_cues, ...) │ ▼ Qwen3-30B-A3B-Instruct-2507 (bf16 MoE) │ ▼ cinematographic image prompt + video prompt (single LLM pass each) VLM (Qwen/Qwen3-VL-4B-Instruct) extracts a structured SceneDescription from the source image — explicit subjects, setting, lighting, and crucially dynamic_candidates / observed_motion_cues so subject motion stays grounded in what the source actually depicts. LLM (Qwen/Qwen3-30B-A3B-Instruct-2507, bf16) composes the final image and video prompts in a single pass each. The hardcoded motion / camera / lens enumeration that previously gated prompt construction is gone. Both models are dropped from VRAM after each batch (clear_gpu()); total resident VRAM during a batch is ~69 GB, sized for an 80 GB card. Diversity & bias mitigation Adversarial inspection of an earlier output batch surfaced three biases (every prompt opened with similar phrases, the same ~5 camera moves recurred, lens choices clustered around 35mm). All three were addressed: Removed anchoring examples from the system prompts. Concrete phrases like "slow dolly-in on a 35mm lens" were replaced with abstract guidance that asks the LLM to range across the full vocabulary. Per-modality prior-prompt feedback loop. The LLM is shown the most recent _PRIOR_WINDOW = 12 prompts from the same batch and instructed to make the next composition demonstrably different across shot framing, camera movement, focal length, lighting, palette, mood, opening sentence, and vocabulary. Backed by a collections.deque(maxlen=12) so the in-context history is bounded — earlier we saw CUDA OOM

Read the release notes on GitHub ↗

Taosis indexes announcements, articles, posts and releases about Bittensor subnets from their own channels and keeps a record of each. The text above is the source’s; the figures are read from the chain by Taosis.