SenseNova U1 and ComfyUI¶

SenseNova U1 is an official image-generation/editing family. smthemex/ComfyUI_SenseNova_U1 is a separate third-party wrapper. Official nodes, official model capabilities, and wrapper capabilities must be attributed independently. Status verified against first-party English/Chinese sources and the wrapper repository on 2026-08-27.

Current Artifact Map¶

Do not attribute every official U1/U1.5 feature to the third-party wrapper. Do not treat the wrapper's quantized path as evidence that the official full-precision model fits the same hardware.

Development History¶

Official model development and third-party wrapper development are connected by integration, not ownership. Represent them as separate subjects/threads with explicit integrates_with links rather than one false predecessor chain.

Hardware Selection¶

The official Chinese FAQ reports approximately 36 GB for full BF16 and requires more than 36 GB VRAM for that path. For constrained devices, it recommends lower/balanced GGUF choices.

The same FAQ warns that a full Q6 path on 8 GB is unsafe and can OOM. Do not convert “GGUF available” into “every GGUF tier fits 8 GB.”

surface: official-nodes-or-third-party-wrapper
official_revision: 022fa663d3f1a3c6822ec0b534e336a991bb7b0b
wrapper_revision: 35f930a349aa38a94100529fd1ec7dd8b31f9ec9
model: U1-or-exact-U1.5-artifact
format: BF16-or-FP8-or-INT8-or-exact-GGUF
adapter: <exact-hash-or-none>
comfyui_revision: <exact-revision>
gpu_and_vram: <exact-device>
steps: <integer>
resolution: <width>x<height>
peak_vram: <measured-value>

Practical Selection¶

• Prefer official registry nodes when the official artifact/runtime contract meets the task. The current official package is ComfyUI-SenseNova-U1 v0.3.0 and requires ComfyUI's v3 node API (comfy_api.latest).
• Install the official package through ComfyUI Manager/Registry or pin its mirror revision. Its README identifies OpenSenseNova/SenseNova-U1/apps/comfyui as source of truth.
• Use the third-party wrapper only when its exact quantization/model path is required and pin both wrapper and model.
• Remove or disable a manually copied legacy ComfyUI_SenseNova_U1 directory before enabling the official package. Both implementations register overlapping node IDs, so keeping both makes load order decide which code runs.
• Use full BF16 only with measured headroom above the documented requirement.
• On 8 GB, start from a documented low/balanced quantization and validate a minimal workflow; do not select full Q6 by name alone.
• For infographic generation, record language, repeated-text failures, background behavior, adapter hash, steps, and resolution.

Training Boundary¶

The official project released full fine-tuning code on 2026-05-21. The official FAQ describes LoRA training code as planned, not equivalent to the already released full fine-tuning path. Re-check this status before claiming LoRA training support because it is temporally unstable.

Local and API surfaces can expose different models, defaults, limits, and output behavior. Keep API results out of local workflow benchmarks unless the exact endpoint is the subject.

Community Evidence Boundary¶

The older eight-step infographic LoRA report is useful project history but is not a current default: its own wrapper documentation now strikes that path through. Exact wrapper issues add context:

• Issue #12 reports an 18.6 GB SenseNova-U1-8B-MoT-8step-Q8_0.gguf edit on a 16 GB NVIDIA card: 4.92 GB GPU allocation was logged and the eight-step prompt took 42.19 seconds, indicating heavy offload/under-utilization. Another reporter measured 37 GB combined GPU/shared memory for a 2K run on RTX 4090 and suggested testing prefetch_count=0. These are workflow-specific observations, not a capacity guarantee.
• Issue #17 reports burned-out results for Interleaved/Infographic fine-tunes under Q4 while the same reporter found Q6 acceptable. Preserve the exact GGUF quantization before diagnosing adapter quality.
• A community U1.5 GGUF model card exists, but it is not an official model card and cannot override the official >36 GB BF16/8 GB Q6 warning.

Gotchas¶

• Issue: Calling the third-party wrapper “official SenseNova ComfyUI” -> Fix: name the repository owner and keep official registry nodes separate.
• Issue: Installing the official v0.3.0 package beside the legacy third-party directory -> Fix: disable/remove the legacy copy first; duplicate node IDs make the active implementation scan-order dependent.
• Issue: Loading official v0.3.0 on an older ComfyUI v1-node runtime -> Fix: update to a build exposing comfy_api.latest or keep the pinned compatible third-party surface.
• Issue: Assuming Q6 fits an 8 GB GPU -> Fix: follow the official warning and select/test a lower documented quantization.
• Issue: Recommending the old eight-step LoRA because a historical post did -> Fix: mark it superseded/struck-through and verify the current adapter path.
• Issue: Claiming official LoRA training from released full fine-tuning code -> Fix: treat them as distinct capabilities and re-check the FAQ.
• Issue: Diagnosing infographic quality without the GGUF tier -> Fix: record Q4/Q6/Q8 and adapter revision; issue #17 reports materially different results between Q4 and Q6.

Temporal Status¶

• Current: official U1 repository and registry nodes; current U1.5/quantization wrapper surface; released full fine-tuning code.
• Superseded but retained: historical wrapper eight-step infographic LoRA instructions now struck through.
• Planned/unstable: official LoRA training-code status.
• Unsafe generalization: full Q6 on 8 GB.

Agent Brief¶

First resolve official v0.3.0 nodes versus the smthemex wrapper, then U1 versus exact U1.5 artifact and quantization. Pin both repositories when the wrapper is used, and never load both node implementations together. Treat historical adapter instructions as superseded when current docs strike them through. Preserve official-versus-community attribution and never promise 8 GB compatibility without an exact measured configuration.

Sources¶

• Official repository: https://github.com/OpenSenseNova/SenseNova-U1
• Official Chinese FAQ: https://github.com/OpenSenseNova/SenseNova-U1/blob/main/docs/FAQ_CN.md
• Official ComfyUI registry mirror: https://github.com/OpenSenseNova/ComfyUI-SenseNova-U1
• Official ComfyUI v0.3.0 metadata: https://github.com/OpenSenseNova/ComfyUI-SenseNova-U1/blob/main/pyproject.toml
• Third-party wrapper: https://github.com/smthemex/ComfyUI_SenseNova_U1
• Official project surface: https://www.sensenova.cn/
• Paper: https://arxiv.org/abs/2605.12500
• Chinese VRAM/offload report: https://github.com/smthemex/ComfyUI_SenseNova_U1/issues/12
• Q4 versus Q6 artifact report: https://github.com/smthemex/ComfyUI_SenseNova_U1/issues/17