You are a research analyst. I am tuning a ComfyUI text-to-image workflow for 'HiDream O1 Turbo' (checkpoint hidream_o1_image_dev_2604_fp16.safetensors). Graph: CheckpointLoaderSimple -> ModelNoiseScale (node 124, input noise_scale) -> SamplerCustom (node 108, inputs cfg + noise_seed; sampler node 230; sigmas node 112 with steps+scheduler) -> VAEDecode. Measured on this exact setup (1024x576): - lcm, normal, noise_scale=8.0, cfg=0.0 -> flat cartoon - lcm, normal, noise_scale=8.0, cfg=1.0 -> clean photorealism - lcm, normal, noise_scale=8.0, cfg=5.0 -> two-tone stencil - lcm, normal, noise_scale=1.0, cfg=1.0 -> pure noise - steps 8 vs 28 at cfg=1.0 -> near-identical, both photorealistic Answer concisely: 1. Is LCM actually the intended sampler for HiDream O1 Turbo, or just what the bundled workflow shipped with? What sampler do HiDream O1 official docs/examples recommend? 2. Why would cfg=0.0 give a cartoon and cfg=1.0 give photorealism in SamplerCustom? Confirm cfg semantics in SamplerCustom vs KSampler. 3. What steps/cfg/sampler does HiDream-I1 (the base family, which used a guidance-distilled dev variant) officially recommend? 4. Is ModelNoiseScale/noise_scale=8.0 a documented HiDream parameter? Documented default and range? Under 400 words. Be specific. If unsure, say so explicitly rather than guessing.