Media Gen Roadmap

Diffusion video/image generation is memory-bandwidth-bound. Measured community scaling on Wan 2.2 vs an RTX 3090 (936 GB/s): 320×320 lands at ~36% of the 3090, 512×512 at ~19%, 640×640 at ~11%. That collapse is an unresolved gfx1151 ROCm kernel inefficiency (TheRock #2591) on top of the raw bandwidth gap. Video works on this box; it is not fast, and it degrades hard above 640px.

The genuine hardware win is fit: no consumer NVIDIA card holds unquantized Wan 2.2 14B. This machine does — that's a batch-mode capability, not a speed one.

• Flux.2 [dev] 32B — best quality available, but BFL non-commercial license for the weights (paid license required in a commercial product) and ~32 GB at Q8. ComfyUI-only; stable-diffusion.cpp has no Flux.2 support yet. License-excluded for any monetized path.
• Flux.1 [dev] — same BFL non-commercial weights; superseded by Z-Image/Qwen-2512 on quality-per-watt anyway. Personal/lab use only.
• F5-TTS / CosyVoice (cloning) — look open, but both carry CC-BY-NC-4.0 model weights (Emilia / WenetSpeech4TTS training data). Banned for monetized use. Chatterbox is the only SOTA-quality cloner with permissive code and weights.
• HunyuanVideo — 35–50 min per 5 s clip on this bandwidth class; strictly worse than Wan on this hardware.
• MiniMax H3 (video+audio joint DiT, Apache 2.0, merged into ComfyUI core Aug 2026) — most active Sep-2026 ecosystem, 19.6–31.7 GB INT8 fits easily, but no Strix Halo perf datapoint yet. Watch-list.
• Meta Muse Image — API-only, no local weights. Out of scope for an on-box build.
• SD.Next / FastSDCPU — smaller communities, fewer gfx1151 datapoints than sd.cpp or the kyuz0 ComfyUI toolboxes. Not worth the deviation.

The 124 GiB GTT pool is already claimed by LLM serving (up to 96–113 GB resident). Only three media workloads are exempt from swap-out:

• Kokoro (~1–3 GB) and whisper turbo (~1.6 GB) — always resident, zero conflict.
• Z-Image-Turbo (~12 GB with encoder) — co-resident whenever ~15 GB of GTT headroom exists (fine alongside Qwen3.8-27B; conflicts only with the 100 GB+ MoE models resident).

Everything else (Qwen-Image full precision, Wan 14B, Flux.2) requires unloading the served LLMs first. Rule: any media job >16 GB swaps the LLMs out and restores after. Treat video as an overnight batch mode, not an interactive one.

1. Phase 0 — zero risk, co-resident: whisper.cpp Vulkan build + large-v3-turbo; Kokoro CPU service. Verify with a 10-min transcription RTF and a 5-min narration render.
2. Phase 1 — low risk: stable-diffusion.cpp v1.0.1 Vulkan build; Z-Image-Turbo GGUF Q8 (7.2 GB) + Qwen3-4B encoder. Bench target ~10 s/image.
3. Phase 2 — medium complexity, LLM swap required: ComfyUI via the kyuz0 Fedora toolbox (closest match to Fedora 44); TheRock multi-arch wheels with [device-gfx1151] extras; the established env set (HSA_OVERRIDE_GFX_VERSION=11.5.1, HSA_ENABLE_SDMA=0, HSA_USE_SVM=0, --disable-mmap, attention_mode=sdpa, cudnn.enabled=false). Then Qwen-Image-2512 Q8 and Wan 2.2 TI2V-5B.
4. Phase 3 — on demand: Chatterbox for voice cloning, watermark-on.

• Images: Qwen-Image-2512 produces black images on k-quant GGUFs (sd.cpp issue #1385, activation dequant overflow) — stay on Q8 and lead with Z-Image.
• Video: the gfx1151 ROCm kernel inefficiency (TheRock #2591) has no fix timeline; budget 10–30 min/clip and cap resolution at 480p for the 14B.
• Audio: the cloning-license trap above; and Chatterbox embeds a PerTh neural watermark in every output by default — a disclosure fact, not a defect.

ComfyUI 4x-UltraSharp (64 MB, models/upscale_models/) — verified live: took the blurry Z-Image 1280×720 output to 5120×2880 in under a minute through the ComfyUI API (LoadImage → UpscaleModelLoader → ImageUpscaleWithModel → SaveImage). This is the standard production path: generate at the model's native resolution, then neural-upscale — native 1080p generation from small models collapses, but upscaled output is sharp. RealESRGAN x4plus also staged. sd.cpp supports --upscale-model on the same files for the non-ComfyUI lane.

All ran co-resident with the served LLMs — llama-swap never stopped. Wan at small resolution ran 4× better than the eval doc's 720p projections. Wiring added: ComfyUI-GGUF custom node, gguf dep, extra_model_paths.yaml, model symlinks. Known gap: espeak-ng (installed 2026-09-10) was missing during the test — Kokoro skipped out-of-vocabulary words; fixed after. Full repro: research/mediagen-live-tests.md.

Full-resolution video battery · 2026-09-11 · complete

Hi-res battery on sd.cpp vid_gen (Vulkan/RADV), creative prompts, GPU decode, 27B LLM evicted during renders (unified-memory headroom). LTX-2.3 dev Q8_0 (21 GB), distilled Q4, and Wan 2.2 TI2V-5B.

Daily-driver video pipeline (operator pick, 2026-09-11): Wan 2.2 A14B + Lightning LoRA — 4 steps (2 high + 2 low), cfg 1.0, 832×480/33f: sampling 287.6 s (~4.8 min) vs the 5B's 418.5 s — the bigger model is also the faster one, and visibly more photoreal (fur micro-texture, true bokeh; the 5B control came out stylized/washed at 20 steps). Wall 8:42 incl. decode. Honest caveats: 14B caps at 832×480 on this box; cfg 1.0 means the negative prompt is structurally inert (official Lightning recipe); LoRA's I64 alpha scalars required an F32 rewrite for sd.cpp (fix documented). A/B log: research/wan-lightning-ab.md.

Reliable ceiling (LTX path): 1920×1088 / 33 f / 24 fps (~1.4 s clips), zero workarounds, GPU decode. During debugging a regression was root-caused to the amdgpu.lockup_timeout=600 kernel param (it hung every large video-VAE decode on RADV/gfx1151); it was removed — do not re-add it to this box's cmdline. Full evidence trail (6 rounds): research/gfx1151-sdcpp-ltx-debug.md.

Photoreal image battery · 2026-09-11 · complete

SDXL-class photoreal models (uncensored-permissive tier) via sd.cpp, shared SDXL components (clip_l + clip_g + VAE), euler, 28–30 steps, cfg 6–6.5, FA. Outputs delivered to the operator's Telegram — not published on this site (governance: uncensored-tier outputs stay off the live docs).

~55 s per 1080p-class image, both models production-usable, no upscale required. ~1.6 s/it, resolution-insensitive in this range. Known non-fatals: recurring Vulkan pinned-mem alloc warning during VAE decode (decodes fine, 8–10 s) and an auto-fit "no GPU devices" warning on some runs (renders still execute on Vulkan). Battery log: research/uncensored-battery.md.

Uncensored-capability proof · 2026-09-11 · proved

Capability test of the uncensored-permissive tier: renders of prompt classes that filtered services block, judged on coherence (uncensored models don't refuse — the claim is that anatomy/structure hold where filtered-trained models mangle). Outputs delivered to the operator's Telegram only, per governance — verdicts only here.

Operational finding: both models self-interpret toward modesty unless the prompt explicitly names the fine-art nude tradition — precise prompting is required. Test surface, hard lines: adults only, no real-person likeness, no non-consent framing. Log: research/uncensored-proof.md.