Qwen Image 2512 · 360 Panorama: mechanisms and practical controls

Start with a scene description and the retained standard configuration: a 2048×1024 canvas, batch size 1, panorama strength 1 and fast mode disabled. The selected sampling settings are 50 steps, CFG 4, Euler, the simple scheduler and denoise 1. These are saved settings; the pack supplies no measured quality comparison.

The exposed image output is output:60, named Sweet Tea Preview Image. VAEDecode 271 supplies IMAGE slot 0 to SweetTeaPreviewImage 60 through its images input. The interface describes node 60 as a no-permanent-save image transport. No permanent filename or export destination is established, so use the client's available output handling if you need to retain the image.

The recipe's supported task classification here is generate_image: it generates from text and an empty latent. The retained compiled operation has no capability or mechanism-rule bindings; the graph and interface provide the configuration facts explained below.

The graph selects these exact files in the indicated ComfyUI model categories:

Category and filename	Loader and configuration
diffusion_models/qwen_image_2512_fp8_e4m3fn.safetensors	UNETLoader 276, unet_name; weight_dtype=default
text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors	CLIPLoader 280, clip_name; type=qwen_image, device=default
vae/qwen_image_vae.safetensors	VAELoader 289, vae_name
loras/qwen-360-diffusion-2512-int8-bf16-v2.safetensors	LoraLoaderModelOnly 275, lora_name
loras/Qwen-Image-2512-Lightning-4steps-V1.0-fp32.safetensors	LoraLoaderModelOnly 277, lora_name

Preparation also lists tea-time-labs/sweet-tea-nodes. All six components are marked required, including Lightning. The graph selects Lightning's model output only in fast mode, but the pack does not establish whether a runtime can omit the inactive branch's file. Prepare the full listed set rather than assuming that switching modes removes an installation requirement.

The base diffusion model and text encoder have unresolved Resource identities. Their filenames remain explicit requirements. The other prerequisite Resource links identify components, not executable fallback recipes. A resolved identity does not establish installation in the current runtime. Model provider versions are recorded as main, with no pinned checksums.

The pack supplies no minimum ComfyUI version, supported platform, scoped GPU/RAM recommendation or comparable run duration. Node version properties do not establish a minimum compatible installation. Preparation metadata labels a 1024×1024 workload, while both graph and interface save 2048×1024; use the latter as the configuration start point, without treating the preparation workload as a measured profile.

PrimitiveStringMultiline 266 supplies equirectangular 360 image, and node 269 supplies the scene. StringConcatenate 267 receives them as string_a and string_b, with delimiter saved as an empty string. The comma and trailing space therefore belong to the prefix itself. Node 267's STRING slot 0 feeds CLIPTextEncode 279's connected text input. CLIPTextEncode 281 encodes the negative prompt separately; both encoders receive clip from CLIPLoader 280.

UNETLoader 276 feeds panorama LoRA 275, saved at strength_model=1. Standard mode selects that result. Fast mode selects the result of Lightning LoRA 277, which takes its model from 275 and also saves strength 1. Fast mode therefore retains the panorama adaptation and adds Lightning. The selected MODEL passes through ModelSamplingAuraFlow 273, with shift=3.1000000000000005, into KSampler 272. The pack does not explain the consequences of changing shift, so it provides no supported shift-tuning recommendation.

EmptySD3LatentImage 282 supplies KSampler's latent_image. There is no source-image input or image encoder in this path. KSampler receives positive conditioning from 279 and negative conditioning from 281, then emits LATENT slot 0. VAEDecode 271 receives it through samples, uses vae from VAELoader 289, and emits the IMAGE sent to preview node 60.

The prefix requests an equirectangular panorama, the LoRA provides the dedicated adaptation, and the canvas sets its proportions. This graph contains no explicit seam-repair stage, and those settings do not establish guaranteed seam continuity. It uses ordinary VAEDecode, with no tiled-decode control.

The turbo_mode port edits PrimitiveBoolean 278's value. That boolean drives model switch 274, steps switch 283 and CFG switch 286, all ComfySwitchNode instances.

Mode	Selected model	Selected steps source	Selected CFG source
false, saved default	Panorama LoRA 275	PrimitiveInt 284, saved 50	PrimitiveFloat 287, saved 4
true	Lightning LoRA 277 after panorama LoRA 275	PrimitiveInt 285, saved 4	PrimitiveFloat 288, saved 1

The toggle selects the current values of those sources; it does not restore them if you previously edited them. Fast mode is four-step only while its active step source remains 4. Its name establishes neither measured speedup nor quality equivalence. Seed, canvas, sampler, scheduler, denoise, prompts and panorama strength remain shared between modes.

Edit the source primitives through the appropriate ports. KSampler 272's steps and cfg, and each switch's switch input, are graph connections. Replacing those connections with scalars would bypass the paired selection. The interface supplies no conditional visibility for these controls, so a displayed Steps or CFG field may belong to the inactive branch.

Keep width and height at 2:1 for the intended equirectangular output. They independently edit EmptySD3LatentImage 282 and default to 2048 and 1024. The interface permits each from 16 to 16384 in increments of 16; it does not enforce the ratio or establish that every permitted size fits memory. batch_size edits the same node, defaults to 1 and permits 1–4096.

panorama_lora_strength edits 275.strength_model, before either branch; turbo_lora_strength edits 277.strength_model and reaches the selected model only in fast mode. Both default to 1 and allow −100 to 100 in increments of 0.01. These schema ranges are not tested quality recommendations, and the pack establishes no best strength for a particular scene.

steps edits 284.value and turbo_steps edits 285.value. The active value must satisfy KSampler's 1–10000 range even though the source primitive controls accept broader integers. cfg edits 287.value and turbo_cfg edits 288.value; KSampler accepts 0–100. Its tooltip describes greater prompt adherence with higher CFG and possible quality degradation at excessive values. That is a qualitative control description, not a measured guarantee for this recipe.

sampler edits 272.sampler_name, saved as euler; scheduler edits 272.scheduler, saved as simple. They select the sampling algorithm and noise-removal schedule. Choose from the interface's listed options; the pack establishes no preferred replacement. denoise edits 272.denoise, saved as 1, within 0–1 in increments of 0.01. Although its generic tooltip mentions retaining initial-image structure, this recipe supplies an empty latent. Lower denoise does not create an image-input path.

negative_prompt edits 281.text. The saved Chinese text requests exclusion of low resolution and quality, malformed limbs and fingers, oversaturation, waxy appearance, undetailed or overly smooth faces, an AI-like appearance, confused composition and blurred or distorted text. This is conditioning, not a guarantee that those defects disappear. trigger_prefix edits 266.value; it is prepended to the scene. If you change it, supply your own punctuation and spacing because node 267 adds no delimiter.

The actions below restore individual saved values. Their effects on Steps and CFG depend on the selected mode; they do not switch modes themselves.

The seed records differ: the interface default is -1 despite its declared minimum of 0; KSampler 272 saves 662848447520565, and its widgets retain a randomize mode. The pack does not establish how the client converts -1 or applies randomization. For comparisons, supply an explicit nonnegative seed through the real seed port and check the effective client setting. A fixed seed alone does not establish identical results across runtimes.

Use the reported failing node to choose a remedy. Model loading, KSampler denoising and VAE decoding are distinct stages. Node 271 is VAEDecode and exposes connected samples and vae, with no tile_size. The retained VAEDecodeTiled source describes a different class: tiling changes decoding, not model loading or denoising. It does not supply an executable adaptation for this graph, so adding a tile_size input to 271 is not a supported remedy.

If panorama structure looks wrong, compare the configuration with the saved prefix, 2:1 dimensions, panorama strength 1 and the selected branch's paired settings. These checks identify known configuration differences. The pack contains no executed examples, measured seam comparison or seam-repair procedure from which to promise a further fix.