1.0 guide · start here · by Holo
Follow the setup steps below, then try a simple prompt. You can leave the detailed controls until later.
Explore an image model by moving through a field.
Drag to find neighbouring images. Stop to let the current image settle. Use the same model, address and controls to revisit a result; use saved PNGs to keep the actual pixels.
Changing resolution, growth history, software or hardware can change the result. Coordinates identify a place in the field, not a guaranteed identical image across machines.
You do not need to find a hidden models folder. You can make your own.
For your first setup, use FLUX.2 Klein 4B and the two companion files linked on the Field website. Follow these steps in order.
Run the Field installer and follow its instructions. Open Field Explorer when installation finishes. Complete activation if prompted, using your purchased licence.
The app and the image model are separate downloads. Installing Field does not install the three model files below.
Press Windows + E to open File Explorer. Choose a location on a drive with enough free space, such as Documents. Right-click an empty area, choose New → Folder, and name it Field Explorer.
Open that folder and create a folder named models. Open models and create two folders inside it: vae and text_encoders. Use these names exactly, including the underscore and the final s in text_encoders.
If you already have a suitable models folder, you can use it instead. You do not need to put it inside the app's installation folder.
Open the Field website's download section. Use its three model download buttons. Wait until each download has finished, then move the files from your Downloads folder to the locations below. Keep the downloaded filenames unchanged.
| Download button | Put the file here |
|---|---|
| FLUX.2 Klein 4B | Directly inside models |
| FLUX.2 Klein VAE | Inside models\vae |
| Qwen3-4B Text Encoder | Inside models\text_encoders |
Your folder should look like this:
flux-2-klein-4b.safetensorsThe main model stays here, outside the two folders below.diffusion_pytorch_model.safetensorsThe FLUX.2 Klein VAE file.qwen_3_4b.safetensorsThe Qwen3-4B text encoder file.Windows may hide the .safetensors extension. You do not need to type it onto the filename. These files do not need to be unzipped. The VAE turns the model's result into a picture; the text encoder reads your prompt. You need both.
In File Explorer, open models — the folder containing the main FLUX file and the two smaller folders. Click the address bar at the top (not the search box), or press Ctrl + L. Press Ctrl + C to copy its location.
For example, it might be C:\Users\YourName\Documents\Field Explorer\models. Yours may be different. Copy your actual address rather than typing this example.
Return to Field Explorer. Open the Model panel. Click the folder input beneath the Rescan and Load buttons, then press Ctrl + V.
Press Rescan. Choose flux-2-klein-4b.safetensors in the Checkpoint dropdown. “Checkpoint” simply means the main image model.
Use the path ending in models, not vae or text_encoders, and not the path to an individual file.
Press Load. Loading large files can take a while, especially the first time. Watch Field's status messages and wait for loading to finish before changing settings.
Field should find the two companion files in their folders. If it asks for one, follow Missing model or companion files below.
Open Reading. In Looking for, type something like a small brass robot on a wooden table, warm window light. Press Apply.
Once an image is visible, drag gently on it to explore nearby variations. Release the mouse and wait for the image to finish. Start with the default controls; you can explore Noise, Variety and States later.
Wait until the image has finished rendering, then use Save with its address for the image alone, or Save framed for a frame with its address. Keep the original PNG if you want to restore its recorded settings later.
Next time, keep your models in the same location. Field remembers the folder you scanned. You do not need to download those files again.
| Model | Companion files / use |
|---|---|
| Z-Image Turbo | Requires its compatible text encoder and autoencoder. Supports Noise and Variety curves. |
| Flux 2 [klein] 4B | Requires the compatible Qwen text encoder and a Flux 2 autoencoder. Supports Noise and Variety curves. Flux 1 autoencoders are not interchangeable. |
| SD 1.5 / SDXL | Supported legacy paths. The new per-step Variety curve is implemented for Z-Image and Flux 2. Noise and Variety curve editing is disabled on other model paths, with stored curves retained. |
Model file size depends on precision and packaging. It is not the total memory required to run the model. Companion files and render size also affect memory use.
Distilled models are designed for short renders. Increasing Settled steps does not guarantee a better picture; compare results with your chosen checkpoint.
Leave the Hugging Face token field empty unless the model explicitly requires access approval and a token. It is not a routine first-run step.
A flat, endless sheet. Every coordinate is a fixed address. Going back to 10821, −33462 with the same setup lets you revisit that part of the field.
Use this when you want to find things and be able to return.
Movement is relative to where you already are. There is no map to return to — each step is a step from here.
Use this when you want to wander and let one thing lead to the next.
Planes are separate sheets. Plane 1 and plane 2 hold entirely different pictures at the same coordinate. New plane gives you a fresh one; Centre returns you to its origin. Both live under Place.
Drag on the picture. Moving frames use the selected size and Moving steps. On release, the app settles at the requested output size. If the climb shows one size, there is no resolution growth.
Everything about how that feels lives under Movement.
Sliders retain their normal behaviour. Click the existing value beside a slider label to edit it in place; press Enter or leave the field to apply it. Escape cancels editing. Values are limited to the supported range and rounded to the nearest supported slider position. The small ↺ icon beside the slider restores its default for the current mode.
Value fields use the displayed units: Guidance, Zoom multiplier, Depth, Step size, Image influence percentage and Form strength. Structure uses the number of composition-size reductions; Rungs uses named choices. Prompt reading values respect the layer lock.
The climb, at the top of that section, draws what the settings below it actually produce: one block per render, the first being what you drag on and the last what you keep. Nowhere else does the program say what the climb will be — it is a consequence of three separate controls, and it is easier to look at than to work out.
The sections themselves can be rearranged. Take one by the brass mark at the left of its heading and drop it where you want it; the order is remembered. Which arrangement is right depends on what you are doing that day.
| Control | What it does |
|---|---|
| Moving steps | How much work each of those frames gets. More steps means a clearer frame and a slower drag. |
| Structure | How small a frame the composition is decided at. Lower means the model commits to fewer, larger forms — and those survive the climb, because growing a picture starts from the smaller one rather than from noise. |
| Spacing | How far apart neighbouring pictures are. Small values give a slow drift through variations; large values cross the field quickly. |
| Step size | How far one drag of the mouse carries you. |
| Rungs | How big a jump each stage of the climb makes. Doubling is the fewest stages and hands the model the most to invent at each one; smaller jumps hand it less and cost a render per stage. Whether that is worth it depends on the picture, so it is a setting rather than a decision. |
Preview speed and image growth are separate choices.
| Preview method | Behaviour |
|---|---|
| Resumable | Runs the opening part of the settled schedule and displays a clean-image estimate. Compatible work at the same address can be reused when settling. Active effect curves may require recomputation. |
| Complete | Finishes a separate short schedule using Moving steps. Settling runs its own schedule; the composition may change. |
An early preview can look smoother than the finished image. It is a different sampling state, even when both are rendered at 1024. A tiny decoder is also an approximation, so preview and full decoding can differ.
| Growth method | Behaviour |
|---|---|
| Latent | Enlarges the model's internal image before refining the next size. The faster growth method. |
| Image space | Decodes the previous size, enlarges the pixels, re-encodes and refines. Experimental; adds work between sizes. Compare it if forms change too much during latent growth. |
Preview and growth method selectors are disabled on legacy model paths that do not use them. Growth method matters only when the render increases in size. Neither method guarantees identical composition across resolutions.
Read the blocks from the moving size to the final size.
Each block represents a render size. With several sizes, each enlargement starts from the preceding result. A single block means the image stays at one resolution.
Structure sets the size at which composition begins. Rungs sets the jumps between sizes. Form strength controls how strongly the previous image is retained as the next size is refined. If outlines double or forms break apart, compare a different strength or growth method.
Draw the curve to control additional noise during each render. Left is the beginning, right is the end; height is strength. A line at the bottom is off. Early changes can affect composition; late changes can affect texture and may leave roughness.
Which stages scales the effect across the climb, from smaller to larger renders. Its eight bars are control points across the climb, not a promise of eight renders. With one render size, the rightmost stage value is used. Outlines mark the weights used by the climb. Open Stage weights beneath The climb to see the table listing the size, weight number and Noise/Variety percentages. During rendering it follows the active sequence; otherwise it shows the configured movement climb.
Where you are, and how the field is laid out around you.
Go to takes a coordinate and travels there exactly — useful for returning to somewhere written down, or read off a saved picture. The number beneath it is the plane; Set moves you to that sheet.
| Control | What it does |
|---|---|
| Zoom | Magnifies the field itself rather than the picture. Zooming in makes neighbouring coordinates more alike, so you drift through fine variations of one idea; zooming out spreads them apart and every step lands somewhere unrelated. It is the scale of the space, not of the image. |
| Depth | A third axis through the field, available in Plane and Relation. In Plane it is an offset added to the existing Zoom-derived depth; zero preserves the previous behaviour. In Relation it controls depth directly. Coordinates and plane stay as they are; depth moves you through them along a direction you cannot drag. Small changes are a good way to find a variation of something you like without losing it. |
What decides the picture at this address.
Six slots hold combinations of four controls: Noise through the render, its Which stages weights, Variety through the render, and variety's own Which stages weights. The panel sits just below the noise controls.
Draw the settings you want, press Set, then 0–5. That number now holds a copy. Slots are labelled empty, saved or in mix. The status confirms the saved slot after the write succeeds. Set then the same number replaces it. Ordinary curve edits and blending do not overwrite saved states. Escape cancels Set.
Click saved numbers to assign Left → Right → Up → Down. The outlined direction shows where the next number will go; click any direction to replace its assignment. You may blend one, two, three or four assignments using any of the six saved states. Unassign clears the assignments without deleting slots.
Drag toward a numbered arm to favour that state; the arm endpoint recalls it exactly. Centre mixes all assigned directions evenly. The percentages show the current weights. The four curves update as you drag, and those blended values feed the moving preview. Release settles the result and grows it if the climb contains larger sizes.
Blend focus concentrates transitions around the selected states. Lower values produce broader mixes; higher values favour the nearest state more strongly. It changes the mixing response, not the model's image resolution.
Slots and assignments are saved in the app's settings folder. Export creates a portable States JSON file; Import replaces the six-slot bank from that file. The slots store effect controls, not prompts, coordinates, images or models. Saved image recipes retain the effective curves without requiring the same slots to remain assigned.
Draw the Variety curve to change conditioning strength during a render. The horizontal axis is progress through each render; height is conditioning perturbation strength. A fixed direction is derived from the field address and scaled before each model step. There is no extra model pass per curve sample. Noise injection stays under the separate noise curve.
Variety's Which stages strip scales this curve from smaller to larger resolution rungs. A zero stage weight disables new variety perturbation for that rung; the image may still contain changes inherited from earlier rungs. Likewise, a zero point on the curve restores the original conditioning at that step, not the image that would have existed without earlier perturbations.
These conditioning curves are implemented for Z-Image and Flux 2. The number of model steps limits the number of curve samples: a one-step preview can only read the beginning. A nonzero curve causes settling to recompute that rung, because its curve sampling changes with the step count.
Explore this image uses the last displayed render as an image reference. Adjust Image influence or move through the field to explore variations. Wait for settling first if you want the finished image, rather than a moving preview, as the reference.
A picture to begin from instead of the field. It is the same operation as growing a render — the model starts from what is there rather than from noise, with some of it painted over. Low influence is a hint about arrangement; high influence makes the output a variation on the picture itself.
The image is fitted to the frame's shape by taking the largest centred piece of it, so nothing is stretched. Something is lost at the edges; the alternative is keeping every pixel and none of the proportions.
Dropping an image on the view does this too — unless the image was saved from here, in which case it takes you back to where it was made instead. The file says which is meant, so there is nothing to choose.
| Control | What it does |
|---|---|
| Canvas | The long edge of the finished picture. |
| Frame | Drag inside it. Across sets how wide the picture is, up and down how tall. The shape snaps to proportions the model can render and both edges land on a multiple of 64, because a size off that grid is quietly changed to one that is not — which would break the promise that a coordinate and a size name exactly one picture. The readout shows the pixels you will get. |
| Settled steps | How much work the finished picture gets. More steps take longer and may change the appearance rather than improve it. Compare short schedules with your model. |
| Guidance | How firmly the model is held to the prompt. Its effect depends on the model. Guidance, its value/reset controls and Avoiding are disabled for the current Flux 2 Klein path. Values are retained when switching models. |
Looking for is your prompt. Avoiding is what to steer away from. Press Apply to make either take effect; typing alone does not, because reading a prompt costs real time and doing it on every keystroke would make the program unusable.
This only appears for Flux 2, and most people should never touch it.
That model does not read its text encoder's output. It takes hidden states from three layers partway through, because the last layer has specialised towards predicting the next word and away from meaning. The defaults — 9, 18 and 27 — are the layers the model was trained against.
The three move together by default, because they are three readings of one depth and the spacing between them is what the model was trained against. Take the lock off to move one alone.
Save framed writes the picture with its address on a plate beneath it. Save with its address writes the picture alone, with the address and every setting hidden inside the file.
Both plain and framed files retain their recipe metadata. Drop one back on the view and the saved address and recorded controls are restored. Image-input workflows also require the original reference image, which is not embedded in the recipe. Keep the original PNG. Screenshots and some sharing services remove embedded settings.
Open Picture → Advanced · diagnostics and enable comparison, then let a Flux render settle. Final · full decoder PNG and Final · tiny decoder PNG export the same finished latent through two different decoders, at the final output dimensions. Both downloads are lossless PNGs with settings.
Comparison adds one tiny decode after settling. It does not add sampling steps. If the tiny decoder is unavailable, the full image remains available and a message explains why comparison failed. Leave comparison off for ordinary use.
The small grid at the corner of the view plots every settled picture you have made on this plane. Hover a dot for a thumbnail and its coordinates; click it to return. Each plane keeps its own trail.
A dot carries the whole setup, not just the position — size, shape, steps, variety, growth, the prompt. Returning to a coordinate under different settings would be a different picture at the same address, which is not returning at all.
The Model panel chooses the files Field uses to generate images. For a new installation, follow Set up Field first.
models folder.Field remembers the folder you scan. You can use your own folder on a drive with enough space; you do not need to locate the default installation folder.
This clears cached working data while keeping the model loaded. It can help free memory between renders. If an image still fails with an out-of-memory message, reduce the image size before trying again.
| Reading | Meaning |
|---|---|
| Position | Where you are. Type into Go to and press Go to travel somewhere exactly. |
| ms / step | How long one step of the model takes. The honest measure of whether a setting is affordable. |
| Latent | The size the model is actually working at, which is smaller than the picture. |
| Decode | full uses the model's main decoder. preview uses the small decoder. If a small decoder is unavailable, Flux and Z-Image use the full decoder, which can slow previews. |
| Model / decode | Time spent sampling and converting the result to pixels. Use this split to distinguish a slow render from a slow decode. |
| VRAM | Two numbers: in use, and held. Held memory is reserved and reusable — a large held figure is not a leak. |
The checkpoint was inspected and is not a kind this runs. The file is fine and on your disk; the model is simply not supported. Flux 1, SD3 and most video models will say this.
No model in the dropdown? Confirm the main FLUX file has finished downloading and is directly inside your models folder. Copy that folder's address into Field again and press Rescan.
Field asks for a text encoder or autoencoder? Check the folder layout above. The text encoder belongs in text_encoders; the FLUX VAE belongs in vae. Make sure you downloaded the VAE linked for FLUX.2 Klein, rather than one for a different model.
If Field still cannot find a companion file, open Model → Companion files. Paste the path to the text encoder file in text encoder, and the path to the VAE file in autoencoder. Press Remember these, then try Load again. To copy a file's path in Windows, hold Shift, right-click the file and choose Copy as path; remove surrounding quotation marks if present when you paste it.
Press Purge GPU. If it happens again, lower Canvas or Moving size. A model that loads and then fails partway is usually being asked for a larger picture than the card has room for, not a broken model.
Moving steps is very low for the number of settled steps. A moving frame runs the opening steps of the settled schedule, and early in a long schedule there is not yet a picture to show. Raise Moving steps, lower Settled steps, or try Preview method → Complete.
Check Moving steps, Settled steps and Preview method first. A preview can be an early estimate rather than the finished sampling state. Turn off Noise and Variety temporarily to isolate their contribution. For Flux, decoder comparison separates the decoding difference from the sampling difference.
Press Apply after editing the prompt. For Flux, Is the prompt landing? can help determine whether changing text changes the model response. Its numeric result is a diagnostic, not a guarantee of prompt accuracy. Restore the default Prompt reading layers if you changed them.
Check ms / step. If it is high at small sizes, the model is probably not on the card — look at the load messages, which say where each piece went.
Use a Windows PC with an NVIDIA GPU with at least 8 GB of VRAM, 16 GB of system RAM, and an SSD with around 40 GB free for the recommended setup. Extra models and saved images need additional space.
VRAM is the graphics card's memory; it is different from your computer's system RAM. Larger images need more VRAM, so meeting the minimum does not mean every image size will fit.
You need internet access for downloads and activation. Periodic licence checks also require a connection. Model loading and generation speed depend on your computer and settings.
Begin with Noise and Variety off. Find a useful Moving size and step count. Then change one control at a time. Save a PNG before a large change, and use States to keep effect combinations you want to blend again.