Nova Reality ZI
Nova Reality ZI is realistic Z-Image checkpoint with great feels of details
All images on example were created with diffusers with custom png tags
Custom Diffusers
Custom sd_embed
Rules
You cannot use the generated images for commercial use if it's not edited (or just turning it to black and white)
You can share images without any restriction if you don't monetize it
Advertising this model to outside is always welcome
Recommend Settings
Sampler: Flow Match Euler
Steps: 9~12
Clip Skip: 1-2
Denoising Strength: 0.65 - 0.8
CFG Scale: 1
Prompt: masterpiece, best quality, amazing quality, 4k, 8k, very aesthetic, high resolution, ultra-detailed, absurdres, scenery, photorealistic, {Prompt}, BREAK, depth of field, photorealistic details
Negative Prompts: anime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured, long body, lowres, bad anatomy, bad hands, missing fingers, extra digits, fewer digits, very displeasing, (worst quality, bad quality:1.2), bad anatomy, sketch, jpeg artifacts, signature, watermark, username, simple background, conjoined, bad ai-generated
Description
Expanding prompt capabilities
Full fp8 >> AIO fp8Full bf16 >> AIO bf16Pruned fp8 >> non-AIO fp8Pruned bf16 >> non-AIO bf16
FAQ
Comments (40)
Images on the showcase are made with the custom sd_embed that allows to do the emphasize and AND
ComfyUI version is available: https://github.com/Faildes/ComfyUI-TC_ADV_ZPrompt
No LoRAs are used on the showcase images
Update:
https://github.com/Faildes/ComfyUI-TC_ADV_ZPrompt
The node is up!
I don't know whether this would work as intended but I'm sure this'll do the thing
@Crody TypeError: 'ZImageTokenizer' (tokenizer) object is not callable
Indeed - how does that node work?
@ev25dqyb595 @vasilypodvornyak988 Fixed the issue
This works the same way as other tokenizer:
Connect the CLIP to the clip (Z-Image model's CLIP), CONDITIONING to the conditioning part like sampler's positive or Negative fold's part (positive / negative >> positive)
use_schedule : This allows to use stepping ( [a:b:ratio] or [a:b:step] which changes the prompt from a to b at desired part )
schedule_steps: steps of the generation
weight_strength: how strong the weight gets (recommend between 1.0 - 3.0)
clamp_min: least amount of the weighting (internal, 0.0 recommended)
clamp_max: most amount of the weighting (internal, higher this is stronger the effect gets, 3.0 - 5.0 recommended)
and_strength: the strength that each prompts on AND are (0.6 recommended)
base_bias: how hard the original (non-AND prompt) generation is, this is set because of the generation between non-AND and AND generations' difference (4.0 recommended)
@Salciano AND (caps and) was the non-natural prompting method which do both left and right at the same amount (it prevents one appearing too much like "red hair, blue hair" being red hair main but using AND like "red hair AND blue hair" do 1:1)
This node emphasize much better compare with those built-in emphasizing method (kind of similar to that of SD1.5/SDXL)
You can see the examples on showcase images
@Crody still cant figure out. I get "TC_ADV_ZPrompt: clip.tokenize() did not return key 'l' for Z-Image" error. May you upload ComfyUI workflow example? Thank's.
@vasilypodvornyak988 I couldn't test it out but it supposed to be work with
Connecting CLIP to the model's encoder and the output to the sampler's prompt (or negative fold, which is the diffusers setup I use)
Please report those issues on the GitHub issue page
@Crody cant login github right now, fresh OS setup, need to restore access first. Here an ComfyUI workflow https://disk.yandex.ru/d/jrVoRbMourd44g I use fp8 scaled z-image-turbo text encoder/clip from here https://civitai.com/models/2169712/z-image-turbo-quantized-for-low-vram ,along your checkpoint. Works well in default approach but that plugin.
@Crody an here is log https://disk.yandex.ru/i/3FSw_uEyd1QVFQ
@vasilypodvornyak988 I reviewed the workflow and seems it's our fault for the encoding glitch
Fixed with the latest update
Besides that, it should not be passed on negative for the KSampler: it should go only on positive
If you want to use the negative prompt, try following below workflow
https://huggingface.co/Chattiori/Workflows/blob/main/novazimage_aio_workflow.json
@Crody thank you. I'll try again now.
@Crody "C_ADV_ZPrompt: cannot find callable HF tokenizer at clip.tokenizer.tokenizer.tokenizer" )))) Nope. New issue. Maybe all that issues is cause of my torch 2.11 dev installation? Anyway - Log: https://disk.yandex.ru/i/3FSw_uEyd1QVFQ and workflow https://disk.yandex.ru/d/NrdRKGiGtcfBiA
@vasilypodvornyak988 Probably fixed with the newest version
The workflow looks not bad but I suggest saving the images
@Crody probably i just get wrong text encoder/clip for your model. Or that is torch 2.11 Dev and ComfyUI issue. But it work in default way... =(
TC_ADV_ZPrompt: cannot find callable HF tokenizer. Expected clip.tokenizer.tokenizer.tokenizer (ZImageTokenizer->Qwen3Tokenizer->Qwen2Tokenizer).
log:
2025-12-30T11:48:00.992784 - Traceback (most recent call last):
File "/media/ai/comfyui/execution.py", line 516, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/media/ai/comfyui/execution.py", line 330, in get_output_data
return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/media/ai/comfyui/execution.py", line 304, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "/media/ai/comfyui/execution.py", line 292, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "/media/ai/comfyui/custom_nodes/comfyui-zprompt/nodes.py", line 42, in encode
cond = advanced_zprompt_encode(
^^^^^^^^^^^^^^^^^^^^^^^^
File "/media/ai/comfyui/custom_nodes/comfyui-zprompt/tc_adv_zprompt_encode.py", line 631, in advanced_zprompt_encode
h, p = encodewith_AND_safe(
^^^^^^^^^^^^^^^^^^^^^^
File "/media/ai/comfyui/custom_nodes/comfyui-zprompt/tc_adv_zprompt_encode.py", line 537, in encodewith_AND_safe
base_h, base_p = encodesingle_zimage(
^^^^^^^^^^^^^^^^^^^^^^
File "/media/ai/comfyui/custom_nodes/comfyui-zprompt/tc_adv_zprompt_encode.py", line 476, in encodesingle_zimage
hf_tok = unwraphf_tokenizer_from_zimage_tokenizer(z_tok)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/media/ai/comfyui/custom_nodes/comfyui-zprompt/tc_adv_zprompt_encode.py", line 445, in unwraphf_tokenizer_from_zimage_tokenizer
raise RuntimeError(
RuntimeError: TC_ADV_ZPrompt: cannot find callable HF tokenizer. Expected clip.tokenizer.tokenizer.tokenizer (ZImageTokenizer->Qwen3Tokenizer->Qwen2Tokenizer).
@Crody also i use Intel B580 GPU and torch without CUDA... on Linux platform. Maybe that is a point... cudaless... Any way - i'll try full precision clip, when download that.
U.P.: Nevermind - same output with full clip.
@vasilypodvornyak988 How about now?
@Crody C_ADV_ZPrompt: cannot find callable HF tokenizer. Expected clip.tokenizer.tokenizer.tokenizer (ZImageTokenizer->Qwen3Tokenizer->Qwen2Tokenizer).
log https://disk.yandex.ru/i/3FSw_uEyd1QVFQ
Get a break, man. You need more feedback, not just from me. I will restore github account, after New Year and will send my feedback there. Celebrate. Cheers!
What in the bloody ell is a Zimage Turbo
@Crody Oh that makes more sense, thanks
from what i know z-image is not trained with sdxl quality enhancers such as masterpiece etc.. and why would anyone willingly add them and start bloating the generations with massive negative and positive nonsense word that dont make sense for an natural language models that has its own llm text encoder that interprets prompt. Don't get me wrong i think the model looks great, but if this keeps going this will end like pony models were you can't generate without score tags, and that's regressing AI not pushing it forward.
idk much about the prompt engineering a T2I model. but I have been playing around with keywords with this model and all I have to say is word choice matters. the difference is subtle and most often not needed but if you are in rush and dont want to prompt using a verbose paradigm, you can still use tag style prompting to get a decent result. just check my image posts. I have natural language, tag, and hybrid prompts. The text encoder really doesnt care because the LLM has seen all three styles before.
@PlayAI yeah sure i know tag style/booru tag prompts work, but i mean in general the masterpiece, best quality etc. all the nonsense tags, it just stupid to train new loras and models with tags that makes no difference
Words matter, but in my testing good quality tags matter very little, as ZIT is already biased to perform the best it can, as all models should.
ZIT has a no-nonsense approach. That is exactly why it performs so well. So don't waste your precious tasty tokens doing autistic things every model should do by default, just like ZIT does. Use (bad) quality tokens, if you actually want bad quality, struggle against the model or test its limits or something like that xD
Remember each token gradually decreases performance. Good quality and detail tokens should be used when you actually need them for something specific, since ZIT is a bit generic and so is its English.
Hi! Does anyone know what the difference between this and vanilla ZIT is? Thanks!
Forgive my ignorance, but what is the difference between these 4 for a user?
Full fp8 >> AIO fp8
Full bf16 >> AIO bf16
Pruned fp8 >> non-AIO fp8
Pruned bf16 >> non-AIO bf16
The only difference they have would be the data type and components so their usability are exact same
AIO (All In One, Full) includes custom Qwen3 encoder and VAE in the file so you don't have to install extra files
fp8 is the lighter data type for the checkpoints
bf16 models exists for model merging and systems that is not supported to the fp8
I'd recommend using Full fp8 model for the best uses
We need an anime Z-Image fine-tune 🙏
So I tried in Forge Neo, but comes up as solid black images. I saw in a different comment about getting text encoders and the like, but I'm unfamiliar with this.
What would I need to do to get this checkpoint to work? (I'm using full model fp8 if that matters).
Are you sure using the Euler sampling method and the Normal schedule type?
@masterdesu464 I did use Euler A and Automatic.
But I'll check later. I'll make another reply if it doesn't work.
try Clipskip=2 (or 3)
why are nipplez every looks like ufos`? :þ
Great Model! Thanks!
Fails to load in SD Next.
It complains that it is not a full model (missing weights apparently), doesn't matter which file I download.
You don't load these as checkpoints in the SDXL sense in SDNext. You have to load the correct diffuser (for example Z-Image Turbo, the base one not one of the quantized branches) and then load this as a UNET. This also means you put the checkpoint in the models/unet folder, NOT the models/stable-diffusion folder.
Hi, same mistake as previous before you fix it for Easy Diffusion: "Error: Could not load the stable-diffusion model! Reason: Error while deserializing header: InvalidHeaderDeserialization"
why most of these checkpoints are overwhelmed with female photos? is it not capable of creating male figures?
use male loras with the checkpoints, and put terms like women, girls, femme, lady etc in negative prompt, set them to :1.4 in forge and go up to about :1.6 in comfyui, then add stuff like (man, male, masculine:1.5), in positive at the very top of the prompt, and you should have far better luck, also, in forge you can use controlnet to edge the checkpoint to adhere to the male figures, and in comfyui use a load image with an image to text node as a reference and link that in a multi prompt near top for same effect. :)
I normally use style prompt node in comfyui as it has 5 prompt combo on it, so i can insert little tweaks above and below main prompt for my body doubles plus add ons. :).
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.