MOSS Image T2I + LF Block Cache — v1.0

If you enjoy and benefit from my work, please support its continued development using the “Tip” button on my Civitai profile or by making a voluntary donation in USDT.

• BNB Smart Chain: 0xdCf9846A1B0F2fe32E19dC9CFdDC5194CC13D108

• TON: UQCeukjBulZw9gni1jv8NLRgtmgHDQ6mUiS2qf3716XobIxA

• TRON / TRC20: THLU3yu8VpoPqP4bMLCp86FRhoCweJ2Qz4

Please send only USDT using the matching network. Thank you!

My NSFW channel: AI Bob Public
My SFW channel: AI DEV LAB
------------------------------------------------------------------------------------------------------

Text-to-image workflow for MOSS-Image SFT with LF MOSS-Image Block Cache, my custom node that speeds up generation by skipping repeated computation between steps. No LoRA, no training, no quality-killing step reduction.

At every step the cache runs the first transformer block and checks how much the result changed since the last full step. If the change is small, the remaining blocks are reused from cache. If not, the step is computed in full. The decision is made separately for cond, uncond and the SLG pass, so CFG and Skip Layer Guidance stay correct.

Features:

• Adaptive block cache for MOSS-Image: decides at every step, no fixed schedule.

• Presets: SAFE / BALANCED / FAST / DRAFT. The chosen preset shows its values and locks them. CUSTOM unlocks everything for manual tuning.

• Works with CFG, Skip Layer Guidance (layer 22) and Sage Attention.

• Works with any sampler and scheduler. Deterministic samplers (euler, dpmpp_2m) give the most reuse.

• First and last steps are always computed in full, so composition and fine details are protected.

• Turns itself off on short schedules (fewer than 10–12 steps), for example with 8-step distillation LoRAs.

• Optional log of every cache decision (verbose).

Quick start:

1. Copy the ComfyUI-LF-MossImageCache folder into ComfyUI/custom_nodes and restart ComfyUI. Then refresh the browser page.

2. Install KJNodes and SageAttention for the Sage Attention node, or bypass that node (Ctrl+B).

3. Open the workflow and select your MOSS model, Qwen3.5-9B text encoder and FLUX.2 VAE. Set CLIP type to moss_image.

4. Choose a cache preset (BALANCED is a good start, the workflow uses FAST), enter your prompt and press Run.

Recommended settings: euler / simple, 25–50 steps, CFG 2.5–3. Keep the cache node right before the KSampler on the MODEL line.

How to pick a preset: set the seed to fixed and generate once with the cache node bypassed and once with it enabled. Then compare. Small text and faces are the first to show the difference. If quality drops, go one preset safer.

Experimental release: the actual speedup depends on your GPU, resolution, sampler and prompt. Preset values are starting points, not per-prompt calibration.

All models download from: Comfy-Org / MOSS-Image
My recommended text encoder: Qwen3.5-9B Uncensored Heretic — ComfyUI